YT YouTube2Text

YouTube2Text → Transcripts → AI Engineering Full Roadmap vs 100-Day Sprint: In-Demand Skills, Job Market Data & Curriculum

AI Engineering Full Roadmap vs 100-Day Sprint: In-Demand Skills, Job Market Data & Curriculum — Transcript

by CampusX · 17,926 words · 2,636 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Good evening again. Let's start the
0:02session. First of all, let me briefly
0:05explain what will happen in today's
0:07session. In today's session, I will
0:11give you a roadmap for AI engineering.
0:14Okay? We will discuss everything. What
0:17is AI engineering? Why should one
0:19become an AI engineer? What are the
0:20possibilities? What is the number of
0:22jobs available? We will discuss all
0:24this, and along with that, I will share
0:26a very nice, lovely roadmap with you. A
0:28very detailed roadmap. I have spent the
0:30last, I guess, four months building
0:32this, and it didn't just take four
0:34months to create. Everything I have
0:37learned in the last two years—if you
0:40notice on CampusX, I started teaching
0:43LangChain in January last year, and
0:45even today we are mostly working on
0:48these topics. So, all that I have
0:51learned about this field in the last
0:53two years, I have tried to put into
0:55this roadmap. My vision with this
0:58roadmap is to provide a perfect guide
1:01for anyone today, whether they are a
1:04beginner or even someone with a bit of
1:06experience, so they can sit down and
1:09prepare for this specific job role at
1:11their own level. Okay? So, this is the
1:15first goal of this session. The second
1:17goal is that I also want to explain a
1:19bit about what we are doing at CampusX.
1:21Many people often scold us, asking what
1:24we are doing. This course is very
1:26important. There is a lot of demand for
1:28it in the market. You aren't even
1:29making it. For example, a lot of people
1:32talk about fine-tuning a lot. We see it
1:35a lot in the comments. And many people
1:37have reached the point of being angry,
1:39asking, "Sir, we are telling you so
1:40much, why aren't you teaching it?" So,
1:42in today's session, you will see that
1:44we are not operating randomly. We have
1:48a vision in mind for what we need to do
1:50in the next six months to a year, and
1:52we are centering all our work around
1:53that as a company. So, you will also
1:56get some clarity on what we are going
1:58to do in the next six months and which
2:00courses you can expect to see. Okay? So
2:02, these are the two main goals of
2:04today's session, and hopefully, I will
2:06be able to achieve both by the end of
2:08the session. Okay? So let's start
2:11today's session and first answer a very
2:13important question: sure, we can read
2:16and watch the AI engineering roadmap,
2:18but why should we study it? What is
2:21special about this job role? Let's talk
2:23about it for a moment. For that, let me
2:26first clearly define what AI
2:27engineering is or who an AI engineer is
2:29. In the market, you will see two or
2:33four very common job roles. Among them,
2:37the oldest one that still exists is the
2:39data scientist. Data analysts are also
2:43quite old. Data engineers are also very
2:47old. And now, in the last three or four
2:49years since ChatGPT arrived, these two
2:51or three new job roles are emerging.
2:54Like an AI engineer, or more recently,
2:57a new job role called FDE is emerging.
3:01Or agentic AI engineer. These are some
3:04of the new job roles coming up. So, we
3:06are going to focus on this particular
3:09job role. Okay? Who is an AI engineer?
3:12In very simple terms, an AI engineer is
3:16someone who develops LLM-based
3:19applications. Okay? An AI engineer is
3:26not the person who trains and builds
3:29LLMs. Those are different. Those are
3:32research scientists working at
3:34Anthropic. Working at OpenAI. Whose job
3:36is to create LLMs. That's a completely
3:38separate job profile. An AI engineer is
3:42basically someone who, if you give them
3:45an LLM via an API or an open-source
3:47model, can take that model and build
3:50applications on top of it. What kind of
3:53applications? Chatbots, RAG
3:57applications, agents, multimodal
4:01applications, LLM-based workflows. So,
4:08building all of this and managing it,
4:10deploying it, looking after its
4:12security, and evaluating it is all the
4:15work an AI engineer does. This is the
4:18description in a nutshell. Now, I would
4:21like to add one more clarification and
4:23a disclaimer that AI engineer is still
4:25not a universally accepted term. So,
4:28what will happen is when you go to look
4:30for a job. On .com or any job portal,
4:32you will see many postings there where
4:35you will see the AI engineer or GenAI
4:38engineer job role. But when you look at
4:41the job responsibilities, you will see
4:43that you have to do all of these things
4:44. But besides this, they expect you to
4:48be good at machine learning, deep
4:50learning, and you should also have a
4:51handle on the concepts of training and
4:53fine-tuning LLMs to some extent. So,
4:56you will see a bit of this variation in
4:58the job market. But mostly, one thing
5:00you will commonly see is that you
5:02should know all these things. You
5:05should know how to build LLM-based
5:06applications. Now, there is so much
5:08variation in the job market because
5:10recruiters are a bit lazy. And during
5:13hiring, they want to hire someone from
5:15whom they can get extra work done in
5:17the future if needed. Meaning, they
5:20gave you an LLM-based application to
5:22build. But tomorrow, if they suddenly
5:25realize, "Oh, we need to build an
5:26XGBoost model on some tabular data."
5:29Then I don't want to hire a new data
5:30scientist for that. I will just give it
5:32to you. Because I had mentioned to you
5:34that you should also know machine
5:35learning. So, this has created a bit of
5:37a gray area. But I feel that over time,
5:40as everything gets a bit more refined.
5:42People will start to get some clarity.
5:44So, over time, it will become a very
5:46clear job role in itself called AI
5:48Engineer, and this person's work will
5:50be very simple. Building LLM-based
5:52applications and managing them
5:54end-to-end. Right? So, this is the
5:57definition of an AI Engineer. I guess I
5:59cleared this up in the first 5 minutes.
6:02Now comes the next question. Why should
6:06I prepare for this job role? So, here I
6:11would like to give my personal opinion.
6:15This is my opinion, and it has remained
6:17the same for the last two-three years,
6:19that I feel this job role will explode
6:21the most. It is already exploding, and
6:24it will grow even more in the future.
6:27Because very few people or very few
6:29companies want to build their own
6:32models. Because it is very difficult,
6:36very costly, and very skill-driven. So,
6:39it is not everyone's cup of tea to
6:40build their own models. On the other
6:44hand, the frontier labs are getting
6:46really good at building models and
6:48intelligence, so they are creating
6:50models for the whole world, meaning
6:53there is no need for everyone to build
6:55models. But yes, in every business,
7:00there are problems where we can fit
7:03these LLMs to solve them. LLMs are
7:07really good at language, really good at
7:12vision, audio, and video. So, these are
7:16the three or four problems that are
7:18quite difficult to solve. And if you go
7:21to build your own models, it takes a
7:22lot of time and many times those models
7:24don't even turn out that good. So if
7:25intelligence is brought and handed to
7:27you, saying, "See, this is the model."
7:29It can understand language well, it can
7:31talk. Then obviously, as a business
7:33owner, as a company, you will be able
7:35to use it to solve a lot of problems in
7:38your company. Like the problem of
7:40customer support. For example, for
7:42years, it has been a very big cost to
7:44the company. You literally keep BPOs
7:47set up. You employ 200 people, and
7:49there is their training, there is
7:51attrition. It’s a big thing to manage
7:55. But if AI agents come, voice AI
7:58agents come, and they become perfect
8:00over time, then just think how much
8:01cost and how much effort a company can
8:04save. Right? So, the same with vision
8:08models—let's take an example. Let’s
8:12assume it's a company like Uber.
8:13Companies like Uber have to pay a lot
8:16of money to their drivers. So over time
8:19, it could happen that they would want
8:20to introduce cars that have vision
8:22models, and they are traversing, and
8:23all the money from the company's cars
8:25is coming to the company. Again, I’m
8:27just telling you what’s coming to my
8:29mind right now. Just what I am trying
8:31to tell you is that because of these
8:33LLMs, we now have the option of
8:35intelligence. And using this
8:37intelligence, we can solve many aspects
8:39of our current business. And that is
8:42why companies will hire AI engineers
8:46who have the capability to use these
8:49LLMs to solve problems. And that is why
8:53I feel this job role will explode, and
8:55it is already exploding. In fact, I
8:57made a video on this topic around 2.5
9:00years ago, saying that this new job
9:02role is emerging. And if you watch that
9:05video even today, most of the things I
9:06said in it are actually true. And if
9:09you want proof, you can go to Google
9:11Trends. In fact, I can go to Google
9:13Trends. And I can show you that this
9:17job role, this term on the internet, is
9:21truly growing. For example, if I type '
9:27AI engineer' and in India, let's not do
9:32the past 24 hours. Let’s do the past
9:36five years. So you can see, this is
9:40what I am talking about. Even until
9:432022, when just ChatGPT had arrived,
9:45people hadn't quite figured out what
9:46was going to happen next. But around
9:502024, this term started increasing more
9:52, and now you can see that even until a
9:55few months ago, it was at its absolute
9:58peak. According to the data so far,
10:00there has been a slight dip, mostly due
10:02to the current economic situation. It
10:05will fluctuate, but yes, there is an
10:07increasing function, which I think will
10:10create more jobs in the future.
10:12Similarly, you can go to. com and you
10:14can search how many job openings there
10:17are around this. And lastly, I would
10:19say that if you don't trust me,
10:21obviously you can say, sir, why should
10:23I trust you? You are just a YouTuber.
10:25Maybe you are saying this to sell your
10:28courses. So I will show you something
10:31authentic that I think if you see, you
10:33will feel that yes, it makes sense. So,
10:36about two or three months ago, I got
10:39this report. The name of this report is
10:43India's AI Workforce Report. And who
10:46published this? Uh, this company is
10:49called Quess. I am pretty sure none of
10:51you would have heard the name of this
10:53company. But it's a big company in
10:55India. In fact, it is so big that it is
10:57a listed company. It's a public company
11:00on the NSE. It is a very big company.
11:02What does it do? This company provides
11:05staffing solutions. Basically, it helps
11:08other companies hire employees. This is
11:12Quess. Okay? Their job is that if other
11:15companies need to hire a candidate,
11:17these people provide them with the
11:19right candidate. And it's a big company
11:22. To date, they claim to have helped
11:24900,000 to 1 million people get jobs in
11:26different companies. And that is why it
11:28is such a big company. Why haven't you
11:30heard about it? Because it is a B2B
11:32company. Okay? It is. We mostly know
11:35more about companies that are B2C. Like
11:37Zomato, we are ordering from them
11:39directly. Daily, so I know who Zomato
11:41is. Who the owner is. But these types
11:43of companies benefit other companies
11:45behind the scenes. That is why it is a
11:46bit hidden. But yes, this company works
11:49in the recruitment domain and the
11:51employment domain. And it publishes
11:54such reports very frequently. So this
11:58report came out in June 2026. Uh,
12:01basically at the end of the first
12:03quarter, and there are two or three key
12:05pointers in this report that I want to
12:07discuss with you. The rest of this
12:09report, we will discuss in great detail
12:10in tomorrow's session. But today, I
12:12will just quote it a little bit. So
12:14here, if you go into this report, it is
12:17around a 44-page report. I have
12:19probably read it twice. The key
12:21pointers here are that they mentioned,
12:24according to our study, we believe
12:27there are currently about 1 million
12:30professionals in India working in AI or
12:33AI-related jobs. And they shared some
12:38great news, so let’s read this line.
12:43I guess this is the CEO of the company.
12:46India's AI workforce has reached an
12:48inflection point with an estimated 920k
12:51AI professionals and this is the main
12:54line. 3.5 lakh active AI-skilled job
12:58postings in the last 90 days. So
13:02basically, what happened is that in the
13:04first quarter of this year, 3.5 lakh
13:07new AI-related jobs have been added to
13:09the market. The market has clearly
13:12achieved scale. The bigger story,
13:15however, is the shift in demand. AI
13:17hiring is no longer concentrated in
13:19data science and machine learning alone
13:21. It is expanding across software
13:23engineering, cloud, cybersecurity, and
13:25various other fields. So this is the
13:28first data point: around 3.5 lakh job
13:30openings have been created in the last
13:323 months. So this is not the data for
13:34the entire year. This is data for 3
13:36months. If you multiply this by four,
13:39even at the same rate, we are seeing
13:42that every quarter we are getting
13:44around three-plus lakh jobs. The second
13:49important thing I found in this report
13:51is a bit further down. It shows which
13:55companies and which domains are posting
13:58these jobs and hiring for these types
14:00of roles. But one critical point I
14:04noticed was that, first of all, this
14:08bracket is seeing the most hiring. If
14:12you are currently in the
14:13three-to-five-year experience range,
14:15that is where the most hiring is
14:16happening. For less than one year, as
14:19in for freshers, the market is very
14:21slow. It is written right here. For one
14:23to two years, it is also okay. For
14:25three to five years, it is very good,
14:26and for more than 10 years, it is a bit
14:28lower for leadership roles. So, this is
14:30an important point. What I want to show
14:34you is which topics and which profiles
14:36are seeing the most emerging jobs. This
14:40is the impact of AI, where the talent
14:42is—there is a discussion about cities
14:43and such here. Yes, look at this: AI
14:46roles with the sharpest demand growth.
14:49Where has the highest growth been
14:51observed? So, do take a look at these
14:52terms once. Agentic AI developer,
14:55basically someone who can develop
14:57agentic AI solutions. AI software
15:00engineer, this is mostly just an AI
15:02engineer. Agentic AI and MCP systems,
15:05GenAI and Agentic AI Engineer, Agentic
15:08AI Architect, RAG and Agentic AI Lead,
15:11GenAI Solution Architect, AI Product
15:14Owner or Product Manager, AI Platform
15:17Engineer, Senior Lead Quality Engineer
15:20GenAI. Mostly you will see that these
15:23are the same things you are studying
15:24here, which an AI engineer does. The
15:27same RAG, the same agents, the same
15:29EVAs, these are all the things being
15:30talked about. And this is where the
15:34most growth has been seen compared to
15:36last year, literally 260%growth, and a
15:38new job role is emerging according to
15:41their own study. It's written here as
15:45FDE, this is where FDE roles are
15:47emerging. And although there are only
15:51100-150 FDEs in the market right now,
15:53over time they are saying that 12,000
15:55to 15,000 more people might be needed
15:57within a year. So in a nutshell, if you
16:00read this report—I will share the
16:01link with you—and tomorrow we will
16:02have a very detailed discussion on this
16:04. So a very big staffing company, Quess
16:07, which has obviously achieved quite a
16:10lot in its field. They are also saying
16:15that this particular job role and
16:17adjacent job roles are going to grow.
16:20Now, what is our outlook until now?
16:23What is our mindset? We hear that
16:26Oracle fired 30,000 people. And
16:30obviously, everyone gets a little
16:32scared and people think that the job
16:33market is finished, and a very negative
16:35mindset is created, and I have heard
16:37this from many students too, that, "Man
16:39, I don't even feel like studying." But
16:42guess what, this is not reality. You
16:44have to study a little bit. You have to
16:46study both sides. You have to study the
16:48negatives too. You have to study the
16:50positives too. Only then will you be
16:52able to come to a balanced mindset.
16:54What is really happening is that
16:56existing traditional job roles will go,
16:59for sure they will go. It is written in
17:01this report as well. If I scroll down a
17:03bit, they have created a section called
17:06"Impact of Future of Work." Here they
17:09have written which job roles are going
17:11to be affected. If you look at the very
17:14top, entry-level engineering, QA, and
17:16support delivery roles have a very high
17:19exposure to automation. This
17:22entry-level engineering is basically
17:23coding and all. So it is true that
17:26coding at the entry level will reduce
17:27significantly and those job roles might
17:29decrease a lot. But they are also
17:33saying that India will create 2.3
17:35million-plus job roles by 2030 and
17:37these will be new job roles. Think
17:40about it yourself, let me give you two
17:41or three examples. If everyone starts
17:44building RAG agentic AI systems, which
17:46people will definitely do. Businesses
17:48will still exist in 2030. They have to
17:50get their work done, so they will build
17:52such systems. Now, to run these systems
17:55, think for yourself, do we need
17:57engineers who know how to run
17:58evaluations? You wouldn't depend on AI
18:01for this. The whole purpose of
18:03evaluation is whether you are able to
18:05correctly evaluate what the AI is doing
18:07or not. You won't have an AI do this
18:09job. Here, you need human
18:11accountability. So these job roles will
18:13exist in 2030. AI security is a very
18:17big thing. Every other day you will be
18:20seeing news that an OpenAI agent hacked
18:22something in some Australian government
18:24website. Or an Anthropic model is doing
18:26something else wrong. So AI security
18:28has to be very, very tight. And this
18:30will become a very important topic. So
18:32in the future, the job role of AI
18:34security experts will exist. Let me
18:36tell you one more thing. AI governance,
18:38basically deciding policies on how an
18:41AI will work. What are the rules and
18:43regulations in different countries and
18:45different states? How can we present
18:47our software according to that? How can
18:49we operate? So there will be a need for
18:52engineers who have an idea about AI
18:54governance. So this new job role will
18:56exist. So what I am trying to say is
18:58that existing traditional job roles
19:01will go, that is already visible. But
19:03it does not mean that there will be
19:05nothing left in the future. Jobs will
19:08remain. There is also a chance that
19:10maybe there won't be as many jobs in
19:122030 as there were during COVID. But
19:14there will be jobs. And man, you only
19:17need one job, right? You just need to
19:19grasp one thing and become so good at
19:21it that you crack that one job. That
19:23should be the mindset. What does it
19:25matter to you whether there are a
19:27million jobs in the market or not? What
19:28is the advantage you have? You have the
19:30advantage of foresight. Foresight means
19:33you can see many things in advance that
19:35other people don't know. I can give you
19:37some examples. If someone is working in
19:39a non-tech field. Has come after doing
19:41an MBA. Is working in sales. Their
19:43day-to-day exposure to AI is very low.
19:45And what is happening in AI on a
19:46day-to-day basis? They don't even read
19:48about it. So they are still largely
19:50shielded. Unaware of what is going to
19:51happen next. But you are someone who is
19:53reading about AI on a daily basis.
19:55Understanding things technically.
19:57Following the daily news. Very neatly
20:00understanding reports like these on
20:02what companies require. Which domains
20:04are hiring. So you have an advantage
20:06that you are able to see what is coming
20:09. So you can prepare better. So here,
20:12it is just a lack of mindset, and that
20:14is why I am taking so much time to
20:16explain this, because mostly what
20:17happens is that as soon as you hear
20:19some news, you are like, "No man, the
20:21future is over, there's nothing left."
20:23That is the wrong move. Be very smart,
20:26read a lot, look at the negatives, look
20:28at the positives, arrive at a balanced
20:30mindset, and then decide what is your
20:32goal, which job role you want to target
20:34. So these AI engineers and related
20:39adjacent things, I feel that by 2030,
20:41it would reach some scale. There will
20:45be jobs because there will be companies
20:46, and companies will need humans. Teams
20:48will not be large. There will be small
20:50teams. But people will require experts.
20:53People will be required around AI
20:55governance, AI security, and AI
20:57application development. So you have to
21:00become that person who has so much
21:02expertise that companies are compelled
21:04to hire you. That should be your goal.
21:06Rather than being very demotivated and
21:09thinking, "Forget it, it's not going to
21:11happen." Then obviously no one can help
21:13you. Okay? So I gave a very long
21:15lecture. But I just wanted to make sure
21:18that before you read the roadmap, you
21:20adopt the mindset that yes, AI
21:22engineering can be a great thing for my
21:25career. Now I would assume that you are
21:29convinced and you want to become an AI
21:31engineer because you feel, "Yeah, this
21:33can be the job role that will get me a
21:36job in the next 6 months to a year." So
21:39now I will quickly tell you how I have
21:43structured our roadmap. Okay? So I have
21:47actually created two roadmaps for AI
21:49engineering. Okay? The first roadmap is
21:54a very comprehensive roadmap in which I
21:58have added everything. Okay? It's very
22:01detailed. And if you go to cover this
22:04properly, it would easily take you 8 to
22:0712 months. Because I have put
22:11everything into it. Nothing has been
22:13missed. This is like the super set. The
22:15superset curriculum of AI engineering.
22:18There is nothing outside of this. All
22:20right? And then, what have I done
22:22second? I have also created an
22:25accelerated curriculum. I have also
22:28created an accelerated road map. Which
22:30is like a compressed version of this
22:32comprehensive road map, which I think
22:35you can easily cover in four to six
22:38months. All right? What is the
22:40trade-off? If you go for this, you will
22:44get 100%of the knowledge, but it would
22:47take more time. And here, you won't get
22:51100%knowledge, but you will get 75 to
22:5480%knowledge, which in my opinion is
22:57enough to crack a job role, or crack an
23:00interview. All right? And the only
23:03advantage of this is that it is faster.
23:05This is faster; you can do it in four
23:07to six months. In fact, I have promised
23:10100 days; I mean, I have designed this
23:12curriculum such that it is divided into
23:14100 days, and I have stated what you
23:16need to do each day. All right? So,
23:18this is the first thing. This is the
23:20structure of the road map. Second,
23:22about resources. We are mostly not
23:25sharing external resources with you
23:28right now. And a big reason for that is
23:32that whatever things are mentioned here
23:34, whatever things you will be asked to
23:35read, all of them are mostly new.
23:38Everything has literally emerged in the
23:39last two to three years. So, there
23:41aren't many trusted resources that
23:42exist in the market. It is not like I
23:44can suggest a book that is very famous
23:46and known to be a good book. Since I
23:48haven't tested any resource, I don't
23:50want to recommend it to you. So, most
23:53of the resources you will see in this
23:55road map will be our own courses from
23:58CampusX. And there will be some things
24:01for which our courses haven't been made
24:03yet. So, I will also give you a
24:05timeline for when we will create those
24:07courses. But over time, I will also add
24:10free external resources here. Because
24:13the road map should have everything.
24:15Why should I only promote my own
24:16courses? I haven't added them right now
24:19only because I found very few resources
24:20that are completely trusted and have
24:22become a benchmark. It is not like
24:24machine learning, which has been going
24:26on for 20-25 years and we know that,
24:27well, "Hands-On Machine Learning" is a
24:29good book. Here, there aren't such
24:31established resources yet. So, this is
24:32the structure of the road map. So in
24:34today's session, I will first explain
24:36the comprehensive one to you in detail.
24:38This will give you a clear picture of
24:40everything you need to study in this
24:41field. What the superset is, and then
24:44we will quickly discuss the accelerated
24:46roadmap—what we will cover and what
24:48we will skip. That will give you
24:51complete clarity on how you should
24:53proceed further. And along the way, I
24:55will also keep showing you which
24:56courses are available. So, you will
24:58find the roadmap on our website. We
25:00have created a new menu here called "AI
25:02Engineering Roadmap." We will be adding
25:05more roadmaps here in the future. An
25:06FDE roadmap is planned. You will be
25:08informed about that in a week or two.
25:10Data scientist roadmaps, and all other
25:12roadmaps will be added here as we go.
25:14For now, you will find the AI
25:15engineering roadmap here. Once you
25:17click on this, a new website will open.
25:21Which we are currently hosting on
25:23Vercel. And this is the website for the
25:25roadmap. First of all, I would like you
25:28all to thank me for building such a
25:30beautiful website. I deserve it. But
25:34anyway, you can see two options at the
25:35top here. The full roadmap, which is
25:38the comprehensive one, and the
25:40accelerated roadmap, which is the
25:41condensed one. Alright? This page has
25:44some added details. How to use it, what
25:47the chapters are? Everything is here.
25:50You can get started right away by
25:52clicking this button: "See the Roadmap.
25:54" So in total, our main roadmap has 16
25:59levels in all. Alright? And you have to
26:03cover things sequentially in a single
26:05order. In total, you will have to
26:08complete around 286 sessions. I am
26:10assuming one session is roughly two
26:12hours. So, it might take you 286 days
26:15to complete this entire roadmap. If you
26:18are putting in two hours per day. If
26:20you put in four hours, just divide it
26:22by two. Alright? So, anyway, I will now
26:24explain level by level what you need to
26:26study in this entire roadmap to become
26:28an AI engineer. Alright? So let's first
26:31talk about level zero. Level zero is
26:34basically your foundations, or what is
26:36called what you need to know beforehand
26:38. What are the prerequisites? So you
26:40need to clear those out first right
26:42here. By the way, one more thing. Uh,
26:43here every level is a page. So, this is
26:46level zero. Uh, this is level one, two,
26:49three. We have given each level a page.
26:52So, navigation is pretty easy. And
26:54apart from that, you will find two or
26:57three things at every level. First, you
27:00will know how many sessions you need to
27:02do at that level and how much of your
27:04complete roadmap is overall finished,
27:07and what things you need to know
27:09beforehand to study that level. What
27:12are the prerequisites? You will find
27:14modules down here. In every module, the
27:16topics and exactly what to cover within
27:18those topics are also mentioned. And if
27:20there are projects, we have shown them
27:23separately in yellow. And lastly, which
27:25courses are available on CampusX to
27:27cover that level, you will find that on
27:30this page. And an interesting thing I
27:33have recently added is that I have
27:34given two numbers for each level. First
27:37, how much weightage that level will
27:39have in interviews, and second, how
27:42much you will use that level on a
27:44day-to-day basis when you are doing the
27:46job. I have given a score to each level
27:48for that. So, this will give you some
27:50idea yourself about what is relevant
27:52for interviews, what is relevant for
27:54the job, and accordingly, you can
27:56personalize it a bit. Okay? So, now
27:58that you have a clear picture. What the
28:00roadmap is like, what it looks like?
28:01Now let's discuss level number zero,
28:03which is like the prerequisite to this
28:05entire roadmap. So here, basically, you
28:08have to be very good at software
28:10engineering fundamentals. Becoming an
28:13AI engineer, I guess, is easiest for a
28:16software developer. Because since you
28:19are not building and training your own
28:21models, machine learning models, deep
28:24learning models. So here, too much math
28:27is not required. Like you would need to
28:30become a data scientist. So truly, only
28:33one prerequisite remains: if you know
28:35how to build and write software well,
28:37you will get a slight advantage in
28:38comparison to someone who has never
28:40written software. So that is why your
28:44level zero is software engineering
28:46fundamentals. I have also mentioned
28:48here that during interviews, its
28:50weightage is very high. Almost 9 out of
28:5210. The reason for that is that many AI
28:55engineering roles have coding rounds
28:57during their screening. Backend
28:59questions are asked. Questions around
29:01APIs are asked. And that is why this
29:03particular level becomes relevant. And
29:06during the job, it is even more
29:07relevant. 10 out of 10, because you
29:09will regularly use software in order to
29:12build LLM-based applications. So both
29:14front-end spaces are a very important
29:16thing that you should learn. Okay? Now,
29:18what all should you learn? I will give
29:20you a quick walkthrough. First of all,
29:22obviously, you should know Python. Okay
29:24? What should you know in Python? You
29:26should be able to write basic programs
29:28in Python. You should have an idea of
29:30basic syntax. This is necessary. After
29:32that, the core data structures, like
29:34lists, tuples, dictionaries—you
29:36should have an idea of these. After
29:40that, you should know how to create
29:42data models like Pydantic and data
29:44classes. There is a video for this on
29:47our YouTube channel as well. You should
29:49have an idea about mutability, scope,
29:51and these things. These are slightly
29:52advanced topics. How to handle files?
29:54How to handle paths? How to handle
29:56environment variables? This is
29:58important. How to write asynchronous
30:01Python code? You should have an idea
30:02about that. Because when you build many
30:05RAG applications or agentic
30:07applications, you make asynchronous
30:09function calls there. So you should
30:11have a bit of an idea about that. How
30:13to handle modules and packages? You
30:15should know unit testing and debugging,
30:16maybe not extensively, but at a basic
30:17level. Because when you ship many
30:20applications, there is a stage called
30:23CI, Continuous Integration. There, you
30:27do testing and all for the application
30:28you have built, so you should know a
30:29few concepts around this. Okay? So you
30:32should know Python well. After that,
30:35you should know quite a bit about APIs.
30:39And to know about APIs, you should know
30:41about these four things: Web, HTTP,
30:43REST, and JSON. I have taught this in
30:47quite some detail in my REST API
30:49playlist, which is on YouTube. The next
30:52thing is that. You should know how to
30:54build APIs with FastAPI. Because any
30:56application you build, whether you are
30:58making a RAG chatbot, an agentic AI
31:00application, or a multimodal
31:01application. Eventually, on the backend
31:03, it will be an API service. So if you
31:06don't know how to build an API service,
31:08you only know one part of the job. For
31:10the other part of the job, you will
31:12rely on someone else. So that is why
31:13you should know this. You should have a
31:16bit of an idea about data, as in
31:18databases and SQL, because you end up
31:20needing it somewhere or other. There is
31:24a good chance that let’s say you want
31:27to give your agentic AI application a
31:29memory feature and you want to store
31:31memories in some database, so there you
31:34need to learn how to set up a database
31:36schema and how to design it. How to
31:40perform basic queries? So, you should
31:42know that. Git and GitHub for sure,
31:44without them you won't be able to
31:46operate because you are going to work
31:48in teams, and there is only one way to
31:49build software in teams, which is Git.
31:52So, you should know things in Git.
31:54After that, I have added one thing
31:56additionally here. I didn't add it in
31:58the beginning but I added it later that
32:01you basically need to know how to use
32:04AI coding tools. Because nowadays,
32:06nobody writes code by hand. Everyone
32:08uses AI coding tools. Now, just saying
32:11in a prompt there like, "Bro, build
32:13this application for me." That is not
32:15the way to work. Spec-driven
32:17development happens in a very
32:18structured way. You should learn that
32:21well. So, I have provided a syllabus
32:23for that here. This is also a playlist
32:25on my YouTube channel for Claude Code.
32:28If you watch that, it is enough. Lastly
32:30, you should have a bit of an idea
32:31about Docker. Because whenever you do
32:33deployment, Docker will be used. So,
32:35your concepts around Docker should be
32:37clear. So these are your prerequisites.
32:40If you have done these, you will have
32:41very little trouble going forward.
32:43Otherwise, you can move forward even
32:44without this. But you will get stuck
32:46somewhere or the other. Okay? We have
32:48added a small milestone project here
32:51where you build your own non-AI REST
32:54API service. The details of what is in
32:56it are added here. You can go through
32:58it once. Talking about courses, you
33:01will find a Python course. In fact, I
33:03will show it to you. This is your
33:05Insiders dashboard. So, you have a
33:09Python course. This is the Python
33:12course. This is like more than
33:14sufficient for whatever is required.
33:16And you will also find Git and GitHub.
33:18I am just quickly telling you which
33:20courses you need to watch. This is the
33:23Git and GitHub course. There is also a
33:25separate SQL course. This is your
33:28end-to-end SQL course. We have also
33:30provided case studies. If you want, you
33:32can read them. We also have a separate
33:34course for Docker. All of this is in
33:37your Insiders dashboard. This is the
33:40Docker course. Advanced FastAPI is
33:42right here. It is coming up next to
33:44this. Flask is not needed. If you are
33:47doing FastAPI, then there is no need to
33:49do Flask. Doing either one of the two
33:50will work. And the Cloud course is also
33:53present in your dashboard along with
33:55the notes. Okay? So, in a nutshell, if
33:58you complete these courses, you don't
34:00need to do everything. In the course as
34:02well, you can skip through certain
34:04sections as mentioned here. If you do
34:06just this much, it is more than enough,
34:09and don't try to become perfect in the
34:11prerequisites. Just set a goal for
34:14yourself that I will spend a maximum of
34:16one month here. And within this one
34:18month, read all of this. As much as you
34:19are able to read. Otherwise, you can
34:21always come back to these things as you
34:23go along. Sir, is it possible to link
34:25direct videos to each written
34:27subsection in each module? That task is
34:29a bit difficult. We do not have access
34:30to links for individual videos. We only
34:32have the course link. There are no
34:34links for individual videos. That is
34:36the problem. Okay? Moving on to level
34:39one. We have finished level zero. So,
34:41your main preparation starts from here.
34:43We assume that you already know the
34:46prerequisites. But let's assume you
34:48have covered them. Now, your main AI
34:50engineering journey starts from here.
34:53So, the very first thing you need to do
34:55is LLM 101. What is LLM 101? Let me
34:59explain it to you. So, most AI
35:01engineers start by interacting directly
35:03with LLM APIs to build chatbots. That
35:06is their step one. I personally feel
35:09that before that, you should spend some
35:12time learning about LLMs. So,
35:15everything is based on LLMs. If you
35:18don't understand the nuances of LLMs
35:20slightly before starting things, your
35:22decision-making around LLMs won't be
35:24that good. So, I have added a step
35:27called LLM 101; the basic goal of this
35:31level is for you to gain a good amount
35:34of knowledge about LLMs. You don't need
35:36the level of knowledge required to
35:38train your own LLMs. But you should
35:41have enough knowledge so that whenever
35:43you need to make any decision around
35:44LLMs, you are aware. For example,
35:47caching is a very important concept;
35:50proper caching can save you a lot of
35:52money. Now, you are an AI engineer who
35:55knows how to make LLM calls. But you
35:57have no idea what proper caching is.
35:59How it works internally. So, an
36:02engineer who knows this will be able to
36:04save more money for their organization.
36:06So, we have covered these small
36:08concepts here in LLM 1001. Its
36:11interview weightage score is eight out
36:13of 10. The score is this high because
36:17many interviewers ask questions around
36:20this topic during the interview round.
36:24Like, questions will be asked about
36:26tokens. You will be asked about the
36:28attention mechanism. Questions are
36:30asked around temperature,
36:31hallucinations, KV cache, and context
36:33limits. Mostly to judge how deep your
36:35understanding is. And here, people
36:38judge you on whether you just make
36:40basic level API calls or if you have
36:42actually studied LLMs more deeply.
36:44Apart from that, the on-the-job
36:46weightage for this level is seven.
36:48Because it will help you make better
36:50decisions regarding cost and latency.
36:53So, in that sense, it is important.
36:55There are 15 modules at this level. Let
36:57me quickly tell you what you need to
36:59cover here. First of all, you need to
37:01understand what an LLM actually is. We
37:04do use them. We make an API call and
37:06get an answer back. But what is it? An
37:08LLM is a next-word predictor; how does
37:10it work? What is auto-regressive
37:13generation? You should know all these
37:15things here. After that, you should
37:17learn about tokenization. You should
37:19learn everything about tokens. Because
37:21in the world of LLMs, everything is
37:23measured in tokens. Latency, input cost
37:26, output cost, everything is measured
37:28in tokens. So you should know a bit
37:30about tokens. That is what you will
37:32study first. After that, you should
37:34have some knowledge of the LLM
37:36architecture. I am not saying you
37:38should go very deep and study
37:40transformer architecture, encoder-only
37:42transformers, or decoder-only
37:44transformers. But you should have a
37:46basic idea that it is a deep learning
37:48neural network where there are layers
37:50stacked on top of each other. There are
37:53different types of layers. There is an
37:54embedding layer. There is positional
37:56encoding. How the attention concept
37:58works. You should have a basic idea of
37:59this. So that if basic questions are
38:02asked, like self-attention is a very
38:04simple concept in itself, where a word
38:06figures out its relationship with all
38:08the other words. So, even if you have
38:11an idea at this basic level and can
38:13explain it in interviews, it's a good
38:14signal to the interviewer that you
38:16understand things in some depth, so you
38:17learn that. After that comes learning a
38:20bit about how LLM training actually
38:23happens. Okay? So there are two things
38:24in this. Pre-training, which is the
38:26main training, and then there is
38:28post-training. You need to study in a
38:31bit of detail how LLMs are aligned and
38:33fine-tuned after training. Okay? You
38:36don't need to go deep into the math.
38:38But at a basic level, what happens?
38:40What is going on in pre-training? What
38:42things are you optimizing? Which
38:44algorithms are you using? Which
38:47optimizer are you using? What scale of
38:49cost do you have to spend? How much
38:51time does it take to train? You should
38:53have a rough idea about all of this.
38:55Okay? Similarly, what happens in
38:57post-training? Why can't we take a
38:59pre-trained model directly to users?
39:01What mistakes will it make? How can we
39:04rectify them? What are the concepts in
39:06alignment? You should know all of this
39:08at a basic level. Once you have learned
39:10how training works, then you should
39:12move to inference, which someone
39:13suggested a little while ago, asking if
39:15there's anything around inference
39:16engineering. So, there is some
39:19discussion here about what happens
39:21behind the scenes in that model when
39:23your prompt goes to an API. How do
39:25prefill versus decode work? How does KV
39:28caching work? You should know where the
39:31costs are incurred during inference.
39:35After that, it is also part of
39:36inference that you read about sampling
39:38and generation parameters, specifically
39:40why there are variations in the output
39:42of an LLM. I am sending something once.
39:45The same question, so an output came.
39:47Next time I asked the same question.
39:48Why is a different output coming? Where
39:50is this difference coming from? Can I
39:52modify it further? Can I make the
39:54output of my LLM a bit more creative?
39:56You need to read about how all of this
39:58is done here. After that, you should
40:00read a little about reasoning models,
40:02which are not very different during
40:04training but behave differently during
40:07post-training and inference. Basically,
40:09they start thinking before answering.
40:11So you should know why these models
40:13behave differently. And what usage can
40:15you extract from this different
40:17behavior? When should reasoning models
40:19be used? When should they not be used?
40:21What is the impact on cost and latency?
40:23When you are using a reasoning model.
40:25You need to study all of this here.
40:27After that, what is a context window?
40:29Why are there limits on the context
40:30window? What happens if you send
40:32something beyond the context window?
40:33You should know this. Because the
40:35context window is a very, very big
40:37discussion when you do AI engineering.
40:39After that, what are the main
40:40capabilities and failure modes of LLMs?
40:42You should know this. Currently, what
40:45is the model landscape? Which are the
40:48proprietary models? Which are the top
40:49companies? Which are the top countries
40:51where models are being built? What is
40:53the scene with open source? You should
40:55know all of this. You should have a
40:58kind of map in your mind. And after
41:00that, this is also an important topic.
41:02Benchmarks and leaderboards, like when
41:03a new model comes out. It outperformed
41:05everyone on this benchmark. What are
41:07those benchmarks? What are those
41:08leaderboards? You should have an
41:10understanding of at least the top ones.
41:12And lastly, how do you choose the right
41:15model for your task? The best model has
41:17arrived. But is the best model the
41:19right model for your work? This is also
41:21an important decision-making process.
41:22You should know how to do that
41:24decision-making. So I personally feel
41:27that if you study this and move forward
41:30, you can do better than any other AI
41:32engineer who hasn't studied all these
41:34things. Because after all, the brain of
41:37whatever applications you build will be
41:40an LLM. So if you are very aware of
41:42LLMs and have deep knowledge, your
41:44decision-making will always be better
41:46compared to someone who hasn't studied
41:48these things. So that is why I added
41:51this, and I didn't start with it when I
41:53made the first version of this roadmap.
41:55But later I felt that I should add this
41:57one. Unfortunately, at this point,
42:00CampusX doesn't have any course where
42:03we have taught you all these things.
42:07But our plan is that within two months
42:09we will give you an LLM 101 course, and
42:11I will teach this course. Currently, I
42:15am teaching you Google SDK starting
42:17next Saturday. So once the AI course
42:20ends, my goal is to complete this for
42:22you in the next one or two months, and
42:24I personally feel that whoever studies
42:26this course will be helped on two
42:28levels. It will help you in your
42:31day-to-day job, and it will also help
42:32you crack interviews. Okay? So yeah,
42:34this was level one, a very important
42:36level in my opinion. Let's move on to
42:38level number two. This is the level
42:40where you formally start your AI
42:42engineering studies. The name of this
42:45topic is Working with LLM APIs. Or you
42:47can also call it LLM orchestration. The
42:50basic idea is that you have an LLM API,
42:52like Anthropic's or OpenAI's. For that
42:55matter, it could also be open-source
42:57models that you are running on your own
42:59machine or your own servers. But
43:01regardless, you have an API, and by
43:03interacting with that API, you build
43:05LLM-based applications or workflows.
43:08This is the starting point. Okay? From
43:10an interview point of view, it is eight
43:12out of ten. Because it is asked in many
43:15take-home assignments and live coding
43:17rounds. And on the job, the weight is
43:19ten. What you study here is what you
43:21will do daily when you are working on
43:22the job. So in that sense, it is very
43:24important. There are around 19 modules
43:26here. I'll quickly tell you what you
43:28need to do. First of all, you will have
43:31to learn how to make an API call. To
43:33any API provider. How do you make an
43:36API call to OpenAI? How to use their
43:39SDK or how to use a library like
43:41LangChain to make an API call and how
43:43to build a simple chatbot where you are
43:45managing conversational state. Okay?
43:48That is what you have to learn here.
43:50After that, you will learn the concept
43:52of streaming. The way the answer prints
43:54on your screen character by character
43:56or word by word, it doesn't happen by
43:58itself. You have to implement that. So
44:00you will learn how to do that here. You
44:03have to learn LangChain because I still
44:05feel it is the most popular library to
44:07build LLM-based systems. So you must
44:10know it. Structured output is also
44:11something that is very important. LLMs
44:13produce text by default. But in many
44:17types of applications, you need to
44:19output some JSON. So that another API
44:22can consume it. So for those kinds of
44:24situations, you will need to learn
44:26structured outputs. Next comes function
44:28/tool calling. Our LLMs produce text by
44:31default. But what if we want to get
44:34some work done by them? That’s where
44:35tool calling comes into the picture,
44:37and you need to study it thoroughly.
44:38Because the fundamental of building
44:40agents starts from here. You will need
44:43to learn how to create your own tools.
44:46How to build multi-step LLM workflows?
44:48Everything discussed so far is
44:50single-step. An input came in, was
44:52given to the LLM, and an output came
44:53out. But it can be more than that. An
44:56input came in and was given to an LLM.
44:57An output came from it. It went to a
44:59second LLM. From that LLM, it went to
45:01three LLMs and then we got an output.
45:03Such flows can exist. Parallel,
45:05sequential, all kinds of workflows can
45:07exist. You will need to learn how to
45:09build them. After that, you need to
45:11learn how you can make multiple
45:13asynchronous calls to LLM APIs. It's
45:16possible you need to make more than one
45:18parallel call. You are building a
45:19coding agent. One agent is reading
45:21files. One agent is creating new files.
45:23One agent is running a command. How do
45:25you do that? You will need to learn
45:27that. You will need to learn the
45:28concepts of concurrency and batching.
45:30After that, how do you make robust API
45:32calls? If the API fails tomorrow. Their
45:36server crashed or there is a problem on
45:38your site, you need to learn how to
45:40handle such errors. After that, you
45:42need to learn the concepts of
45:43observability. Basically, every API
45:46call you make will involve some latency
45:48. It will involve some cost. Logging it
45:50somewhere in tools like LangSmith. You
45:53will need to learn that. And lastly,
45:55you should learn some other
45:56orchestration frameworks such as
45:58LlamaIndex. Although I wouldn't suggest
45:59LlamaIndex that much. Very few
46:01companies are using it. LangGraph is
46:02something you should learn, but again,
46:04that is covered ahead. DSPy is also
46:06something some companies are using. But
46:09I would not recommend it that much. If
46:10you have plenty of time, then do it.
46:12Otherwise, you can leave this. Okay?
46:14And there is a capstone project. Again,
46:17you can go through this and build this
46:19one. For this level, we have three
46:22courses in your dashboard. The first is
46:25GenAI using LangChain, which is
46:27basically a YouTube playlist where you
46:29will find notes, etc., in one place.
46:31Yeah, this one. Okay? And apart from
46:36this, there is a course on Ollama that
46:38teaches you how to do all this same
46:41work with open-source models. And there
46:44is a Gemini-specific course where you
46:47are taught how to do the same things
46:49within the Gemini ecosystem. So if you
46:51do these three courses—well, if you
46:53just do this first one, that’s enough
46:55. But if you want a taste of
46:57open-source models as well, you can do
46:59the Ollama course. I would not
47:01recommend the Gemini course because it
47:03just teaches you the same things within
47:05the Gemini ecosystem, which your
47:06company may not necessarily use. So you
47:08can skip this one. We have mentioned
47:10everything. So yeah, this is the first
47:12proper thing you should learn as an AI
47:15engineer. This is super important. You
47:16cannot skip this. This is not optional.
47:19So, what have we learned up to this
47:20point? We have cleared the basic
47:22prerequisites for software, grasped the
47:25fundamentals of LLMs, and now we have
47:27learned to make basic API calls. We
47:29have learned to build basic LLM
47:31workflows. Now, in this step, we need
47:33to learn prompt engineering next.
47:37Prompt engineering is also something
47:39very important because every time you
47:41build any LLM-based software, you have
47:43to create system prompts, and those
47:45system prompts are very critical. They
47:48significantly change the behavior of
47:50your software. So, writing prompts in
47:52the right way is also an important
47:54thing. Don't mistake prompt engineering
47:56for those YouTube ads that claim
47:58they'll make you a prompt engineer in
48:01one day. That is a separate thing. That
48:03is at the user level, like a chatbot
48:06similar to ChatGPT. How you can get
48:08great outputs by writing prompts in it.
48:11We are not talking about that. We are
48:12talking about developer-side prompt
48:13engineering. Where you have to write
48:15system prompts to build LLM-based
48:17applications. This is different. Its
48:19weightage in interviews is not that
48:21high, however. It is asked very rarely
48:23in interviews. It has on-the-job
48:25weightage because for every application
48:27that you will build, you have to write
48:29prompts at every level. You have also
48:31seen how many system prompts we wrote
48:32in the current Llama-based course. I am
48:35pretty sure you must have written many
48:36prompts in Himanshu’s LLMOps course
48:38as well. Right? In total, there are 11
48:41modules; this course is short. What is
48:44a prompt? You need to understand this.
48:46You need to learn how zero-shot and
48:48few-shot prompting is done. You need to
48:50learn how prompting is done with
48:51reasoning models. You need to learn how
48:54system prompts are designed. How do you
48:56control the output? You need to learn
48:58how to control the output an LLM
48:59provides through prompting. Prompt
49:02reliability and debugging; if there's a
49:03problem with a prompt tomorrow, how do
49:05you debug it—that is important. How
49:07to version your prompts? Over time, you
49:09have changed your system prompt many
49:10times. You changed it 10, 15, or 20
49:11times. Now, if you need to roll back,
49:13but you haven't stored it anywhere. So,
49:15you need to learn how to manage your
49:17prompts. How prompt injection happens
49:19and how to prevent it is also something
49:21you must learn. So these are the things
49:24you need to do, and we have created a
49:27course for you on CampusX called Prompt
49:31Engineering. This course covers all
49:34these things. Okay? So if you watch
49:37this course, you will be able to figure
49:39out all these concepts. Okay? In fact,
49:42there are additional things in the
49:43course as well. But you only need to
49:45cover these specific topics. Prompt
49:46engineering is the next thing. It is a
49:48small thing, but it is important. So
49:50now let's move forward. Now, the
49:53heavyweights come in—the most
49:55significant things to become an AI
49:57engineer, which is RAG. It is still
50:00very important today. Many people will
50:02keep saying that RAG is no longer
50:03useful. RAG is dead. That is not true.
50:06Even now, most of the projects being
50:07built in the AI engineering domain in
50:09the industry are mostly RAG projects.
50:12Because every company wants an internal
50:14chatbot so that their team can chat
50:16over their internal knowledge base of
50:18documents. Everyone has this use case.
50:21And that is why this is super critical.
50:23So, the interview weightage is 10 out
50:25of 10. Everyone I have spoken to who
50:28has cracked jobs has said that the most
50:30questions were asked around RAG. They
50:34were grilled the most on RAG, and the
50:36on-the-job weightage, I would say, is
50:38also 9 out of 10 because most companies
50:40will ask you to set up a RAG project
50:42for them. Okay? In that sense, it is
50:45very important, and it is also our
50:46biggest module in this entire roadmap.
50:48There are around 27 lectures for you.
50:51What do you need to learn? I will tell
50:53you. First, you need to learn how to
50:54build a basic RAG application. For
50:56which you need to learn how to do data
50:58ingestion and document processing. You
50:59will have to learn all of this. The
51:01chunking process—the work of breaking
51:03data to put into a vector database—
51:05you have to learn how it is broken down
51:07and on what basis. You will have to
51:09learn the fundamentals of embeddings.
51:10Because in a vector database, you store
51:13embedding vectors. You will need to
51:15study about similarity and semantic
51:17search. Many interview questions are
51:19asked about this. How do vector
51:20databases work? How are things stored
51:23in them? How are they retrieved? You
51:25will have to learn that. These are all
51:27slightly advanced topics. Sparse and
51:28hybrid retrieval. BM25 and all these
51:30are algorithms. You will have to learn
51:32all of this. Up to this level, you can
51:34build a basic RAG application. After
51:37that, we have covered some slightly
51:39more advanced topics in RAG here. Like
51:42reranking, where you get results from
51:45the retriever. But are they in the
51:48right order, or can they be improved?
51:50Then that part comes under query
51:51transformation. The user gave you a
51:53query. But is that the best way to
51:55fetch things from the vector database?
51:58So, can you make it better? Can you
52:00transform it? You must have seen
52:03grounding and citation, where your
52:05answer says this specific line was
52:07taken from this source. This specific
52:09line was taken from this source. How to
52:11do that? The different variants of RAG
52:13like Self-RAG, CRAG, Adaptive RAG,
52:15Graph RAG, Multimodal RAG—you have to
52:18cover all of this here. And then there
52:20are some frameworks you can use to
52:22build RAG applications. You need to
52:24cover those. So in a nutshell, you can
52:26divide this entire section into two
52:27parts. Building basic RAG applications
52:29and building advanced RAG applications.
52:31Right? And the good thing is that we
52:33have a complete course for this as well
52:35, which Himanshu created, and its
52:37feedback is also very good. Most people
52:39have liked this course. It is a RAG and
52:42advanced RAG course. You will find it
52:44in your dashboard. Right? So this is
52:46the most important module of this
52:49roadmap. This is the bulk of this
52:51course. Of this roadmap. You need to
52:53cover this thoroughly. Even if you
52:56spend a month or a month and a half on
52:57this, if you learn it well, you will be
52:59able to handle most questions
53:00confidently, and the interviewer will
53:02also give you some leeway, thinking,
53:03well, maybe they couldn't answer the
53:05other things, but they were very strong
53:07in RAG. So, I need to get RAG
53:09applications built, and this person is
53:10right for the job. So, this is where
53:12you need to focus. Right? So, level
53:14number four was RAG. Just after RAG
53:17comes the most hyped part of this
53:19roadmap, which is AI agents. Again,
53:23something that is talked about a lot
53:25and is important; the growth of agentic
53:27applications is huge. It has exploded,
53:30and almost everyone is developing
53:32agents. So you will have to focus quite
53:34a bit here as well. RAG and agents are
53:36the core of this entire roadmap. Its
53:38interview weightage score is nine out
53:40of 10. It’s a very hot topic. It’s
53:42quite simple. It is written here as one
53:44of the hottest topics in current job
53:45descriptions. You will see it in every
53:47JD. The on-the-job weightage is a bit
53:49lower. The main reason for that is not
53:51every company builds agents. That is
53:52the reason. Many people build simple
53:55fixed workflows, which are not agents.
53:57People call them agents, but they
53:58aren't actually agents. An agent means
54:00you cannot determine its workflow. It
54:02will vary every time. So if you are
54:04using a simple workflow, like an email
54:07coming in—You give it to an LLM. It
54:10categorizes which team it should go to.
54:12Then this is not an agent. People call
54:14it an agent. But this is not an agent.
54:16This is a simple LLM-based workflow.
54:18Once you understand that distinction,
54:21you will realize that not every company
54:23builds agents. So that is why
54:25on-the-job weightage is low, but it
54:27will increase in the future. Okay?
54:29There are around 22 lectures in this.
54:32What do you need to study? I will tell
54:33you quickly. First, you will need to
54:35learn some theory about what an agent
54:37is. How will you cover the journey from
54:39an LLM to an agent? What are the things
54:41involved? Agents have tools, and they
54:43have memory. It has its own prompt. You
54:46will have to study that. After that,
54:48you will have to learn about types of
54:49agents. ReAct, planning agents,
54:51plan-and-reflect agents. You will need
54:53to learn a little about these different
54:55variants. Then you will have to learn
54:56the concept of tools. Because the
54:58agent's hands and feet are its tools.
55:00MCP is a very famous thing. Basically,
55:03it is a protocol to connect your agent,
55:05your agent's brain—the LLM—and the
55:07tools. We have a very detailed YouTube
55:10channel playlist on this that you can
55:12watch. So you have to study MCP. After
55:15that, memory is super important. LLMs
55:17plus tools plus memory equals an agent.
55:20So memory is super critical. If an
55:22agent does not have memory, it won't be
55:24able to improve over time and will
55:26frustrate you every time. So memory is
55:28something that is very critical to
55:30agents. You should learn that. Then,
55:32multi-agent architecture is a bit of a
55:34complex task, so one agent won't do it.
55:36A team of agents will do it. So, how do
55:38those teams of agents interact with
55:39each other? How does the hand-off
55:41happen? How do protocols like A2A work?
55:43You have to study all of that here. You
55:46should learn LangGraph because that is
55:49the best library right now to build
55:51agents. So, LangGraph fundamentals and
55:54sandbox execution. I added this
55:56separately. I didn't add it in the
55:57beginning. It has become very critical
55:59now. Agents are a bit dangerous because
56:01they have the power to do things. So if
56:03you leave them unchecked, it's very
56:05dangerous. So you generally test them
56:07in a sandbox environment. You give them
56:09a safe environment. A few days ago,
56:12when I taught you Agent Evals, I told
56:14you that we are building a file
56:15management agent whose job is to manage
56:17files. We didn't just leave it directly
56:20on our files there. I gave it a test
56:23folder, and it worked only inside that
56:25test folder. So that test folder was a
56:27sandbox. So whatever mistakes you make,
56:29whatever harm you create, do it all
56:30inside that folder. There is no problem
56:32. Don't do it outside. Once I test that
56:34you are working correctly, then I will
56:37let you out. So, sandbox agents are a
56:39very important concept. A lot of
56:41startups are coming up in this domain.
56:43If you read a bit, in the US, in
56:45Silicon Valley, many startups are
56:46coming up that provide these sandbox
56:48environments. Like, literally there are
56:50libraries. You install them once, and
56:52they can create a sandbox for any type
56:54of agent you have. So, it's a very
56:55interesting thing. After that, agent
56:57reliability, UX, and failure recovery.
56:59Especially for long-running agents.
57:01They are working for very long
57:03durations. Working for 7 days at a time
57:05. Working for a month at a time. So,
57:07there are high chances of failure in
57:08such long-running agents. So, how do
57:10they recover from failure? That is also
57:12a very interesting topic. Recently, you
57:15might have read about some new software
57:17that has arrived. ChatGPT's Tools,
57:20Meta's Muse, Grok. These are all
57:24long-running agents that are working
57:2624/7. So, there are high chances of
57:28failure there. How you make them
57:29recover from those failures is what you
57:31have to study here. And there are many
57:33agent libraries. I would not recommend
57:35that you study everything. I will
57:37recommend two libraries to you. Based
57:38on the current state, learn LangGraph.
57:41It's the best low-level library to
57:43build agents. Because you get the most
57:46control in the process of building your
57:48agent. But it's a bit difficult to
57:50learn because there are many things.
57:52And one is Google's ADK because many
57:54industries and companies are using
57:56Google's products. So they have simply
57:58used Google ADK to build agents. So ADK
58:01is a good option. There are other
58:03libraries besides that. There is CrewAI
58:05, Agno, and there was also some library
58:08from Microsoft. I forgot its name. The
58:10name was AutoGen, perhaps. So whatever
58:12it is, now they are gradually getting a
58:13bit weak. The top two libraries right
58:16now, in my opinion, are LangGraph and
58:19Google ADK. Although you can write
58:21custom Python code to build agents, I
58:23wouldn't recommend that right now. Okay
58:25? So these are the courses that you
58:27need to study at this level. We have a
58:29course on LangGraph. Uh, this is
58:31actually a YouTube playlist, but you
58:33will get notes along with it. This one,
58:36Agentic AI using LangGraph, there is a
58:38CrewAI course too. I guess CrewAI makes
58:41it very easy for you to build
58:43multi-agent architectures. Basically,
58:45you can create a crew of agents. It was
58:48very popular last year. This year there
58:50isn't much talk about this library, so
58:51I don't know. Agno is similarly a
58:53library that helps in building agents
58:55very easily. We have a different course
58:57for MCP. It's a YouTube playlist, this
59:01one. And there are two-three things we
59:04haven't covered in agents yet. So our
59:07focus is on that. Like how to build
59:09multi-agent systems, I haven't taught
59:11that yet. So that is my plan to teach.
59:13I haven't taught Google ADK. So we are
59:15just starting Google ADK from this
59:16Saturday. And I will teach both things
59:18in Google ADK. I will directly teach
59:20you how to build multi-agent systems in
59:22Google ADK. So you will learn
59:23multi-agent systems as well. You will
59:24learn A2A as well, and you will learn
59:26Google ADK as well. So all three things
59:28will be covered together. And this one
59:30course is left, Memory in LLMs. I
59:33haven't taught this in detail. We will
59:34prepare a course on this in the next
59:36one or two months. The curriculum is
59:38finalized. We will just start it. For
59:40now, there is a short YouTube playlist
59:43on my channel where I have taught about
59:45memory. Short-term memory, long-term
59:47memory. Memory is what is the
59:49requirement of a memory in an LLM-based
59:51system. I have covered all of this in
59:52three videos. If you watch that
59:54LangGraph playlist as well, it's a good
59:56start. You can handle interviews. It
59:58might not help as much with building
1:00:00projects. So we will bring a course on
1:00:02this soon, within two to three months.
1:00:05So yeah, this is the bulk, the core of
1:00:09the roadmap: RAG and AI Agents. This is
1:00:13the most critical part. Okay? So, what
1:00:15has the entire journey been like so far
1:00:18? We did foundations, LLM 101, LLM
1:00:21fundamentals, and learned to build
1:00:23basic workflows with LLM APIs. We
1:00:26looked at prompt engineering
1:00:27fundamentals because we need to build
1:00:29system prompts. We looked at RAG and AI
1:00:31agents. As for DSA concepts, this is it
1:00:34. The interview weightage will be
1:00:36slightly higher. That also depends on
1:00:38which company you are interviewing for
1:00:41and at what level. It is asked more at
1:00:43the fresher level. But on the job, the
1:00:45utility of DSA will be almost zero. On
1:00:47the job, you will use DSA concepts next
1:00:50to nothing in AI engineering. So now
1:00:52you know how to build agents. Let’s
1:00:54look at the next thing. The next thing
1:00:55is also important and has recently
1:00:58become popular. Context engineering;
1:01:01basically, your LLM has a context limit
1:01:04. There are many LLMs that have a
1:01:06context limit of 1 million tokens. So,
1:01:08as it starts to fill up. Your agent or
1:01:12your LLM starts to forget things and
1:01:14its performance starts to decline. So,
1:01:18a new field is emerging here called
1:01:20context engineering. Basically, this is
1:01:23someone who manages an agent's or LLM's
1:01:25context in real-time. Right? Take an
1:01:28example. There is a coding agent. In a
1:01:30coding agent, you have many large
1:01:33codebases. There are many files in them
1:01:35. Now, in real-time, you are sending
1:01:37all those files to the LLM and getting
1:01:39it to create new files based on them.
1:01:41So a lot of the context is being
1:01:43occupied by that. If your context limit
1:01:45is breached, your coding agent's
1:01:47performance starts to deteriorate. So,
1:01:49in real-time, what do you need to do?
1:01:53You need to manage what should be kept
1:01:55in the context right now. Which things
1:01:57should be removed and then brought back
1:01:59later. So this entire management is
1:02:01what we study in context engineering.
1:02:03Okay? So this is also kind of important
1:02:05. It's not as important as learning RAG
1:02:08and agents. But it is gradually
1:02:09becoming important. From an interview
1:02:11point of view, it's not that important.
1:02:13You won't see that many questions on it
1:02:15. You might find a few more at the
1:02:16senior level. You won't see it as much
1:02:17at the junior level. But on the job,
1:02:20the weight is good. It's around eight
1:02:22out of ten. Because context management
1:02:25is kind of important. Whenever you are
1:02:27building any kind of LLM-based
1:02:29application, this is kind of important.
1:02:31What do you need to study here? First,
1:02:33you need to understand what a context
1:02:35window is and how to treat it as a
1:02:37resource. After that, you need to
1:02:39understand what things go into a
1:02:40context. What things constitute a
1:02:43context? You need to understand that.
1:02:46After that, how do you compress and
1:02:49compact a context? Suppose your entire
1:02:52context is full, then how will you
1:02:55compact it? How can you summarize it?
1:02:57What are the things that should be
1:02:58summarized? What are the things that
1:02:59shouldn't be? You learn all those
1:03:01techniques here. Memory and state
1:03:03offloading: what things can you put
1:03:05into memory? What things do you need to
1:03:07keep in working memory, as in the
1:03:08context? You have to learn the
1:03:09decision-making process for this here.
1:03:12After that, long context strategies;
1:03:14basically, there is this very important
1:03:17phenomenon. Lost in the middle.
1:03:20Basically, what LLMs do is remember the
1:03:22beginning things. They remember the end
1:03:24things. They forget the things in the
1:03:25middle. And a proper study was
1:03:27conducted on this. And this is a very
1:03:29common problem with LLMs. You have to
1:03:31learn how to handle this kind of
1:03:33situation. Tools themselves emit a lot
1:03:36of context. Their inputs and their
1:03:39outputs are all added to the context.
1:03:42So over time, if your agent is using
1:03:43many tools, all of that keeps getting
1:03:45added to the context. You also need to
1:03:47learn how to manage that. You need to
1:03:50study how to cache context and how to
1:03:52manage it efficiently. And lastly, an
1:03:54important new thing is coming up:
1:03:55harness engineering. This is not
1:03:57directly related to context, but it is
1:04:00an adjacent topic, so I included it
1:04:02here. So harness engineering is
1:04:05basically creating a harness around
1:04:07your LLM in order to solve a given task
1:04:09. For example, if you are evaluating
1:04:13LLMs. Then, to evaluate the LLMs, you
1:04:15called upon an LLM judge. Now, with the
1:04:18help of that LLM judge, you have
1:04:20written code to evaluate the entire LLM
1:04:22. You built a system. So, we call that
1:04:25an eval harness. Or, you took an LLM
1:04:28and built an agent around it, and that
1:04:31agent is using many types of things.
1:04:34It's using tools, using memory. It's
1:04:36using your personal database. So, the
1:04:39custom code you wrote to glue all these
1:04:42pieces together, the code you wrote to
1:04:45connect all these parts, basically
1:04:47became an agent harness. So, harness
1:04:51engineering is something that is
1:04:52becoming very popular right now. These
1:04:54important tools you might have heard of
1:04:56—perhaps at the beginning of the year
1:04:58, a tool called Open Interpreter came
1:05:00out. It was basically an agent harness.
1:05:03There is a tool called Hermes Agent.
1:05:05That is also a harness. In fact, these
1:05:09things like Devin, Muse, and Grok-bots
1:05:11are all agent harnesses around LLMs. So
1:05:14, you will study a bit about these as
1:05:16well. There is no need to go too deep
1:05:18into it for now. Because, from the
1:05:20people I have spoken to who have given
1:05:23interviews, not many questions are
1:05:25asked here. And it's also not needed
1:05:28that much on a day-to-day basis in jobs
1:05:30yet. But yeah, it's a hot term. In that
1:05:33sense, you should know a little about
1:05:34what is happening. Did we conduct a
1:05:36session on harness engineering? No, we
1:05:38probably did one on loop engineering,
1:05:40which is like a subset of harness
1:05:42engineering. But okay, in the future,
1:05:44we can do a session on harness
1:05:45engineering as well. So yeah, context
1:05:48engineering is also kind of important.
1:05:50Not as much as the previous two modules
1:05:52. But you should know about it. Moving
1:05:56on after that. Let's go to module seven
1:05:58. Module seven is also very important,
1:06:01which is basically LLM Ops. This is a
1:06:04completely new field emerging, and it
1:06:06will become very important in the
1:06:07future. This could literally be as
1:06:10important as software testing is. If
1:06:14you remember—well, what would you
1:06:16remember—meaning today there are
1:06:17software engineers and also QA
1:06:19engineers. Basically, who are called
1:06:21software testers. So, a long time ago,
1:06:24when software was written, the software
1:06:26developers themselves used to do the
1:06:27software testing. I am talking about
1:06:30maybe the year 2000. But over time,
1:06:33people realized that writing software
1:06:34is a different specialized job and
1:06:36testing software is a different
1:06:37specialized job. So these two job roles
1:06:41diverged, and today both of them exist.
1:06:45A similar thing will happen in the case
1:06:48of LLMs, where building LLM-based
1:06:50applications and testing them is the
1:06:52same person's job today. But over time,
1:06:56this field will become so specialized
1:06:58that in the future, you will have
1:07:00specialist LLM application testers.
1:07:02That is my prediction. Again, I could
1:07:04be wrong, but this is significant
1:07:06because, until now, we were simply
1:07:08happy making demos, like, "Hey, I built
1:07:10a chatbot.""This is my agent." But many
1:07:13people realized that when they deployed
1:07:15those agents. When people deployed that
1:07:18RAG chatbot, it wasn't performing as
1:07:20well as people expected. So, that's
1:07:23when people realized that just building
1:07:25and launching it directly...Obviously,
1:07:27it's not logical. You have to test it
1:07:29as well. So now, this completely new
1:07:31field has been emerging over the last
1:07:32year or so. Where a lot of work is
1:07:34going on around testing LLM-based
1:07:37applications. Testing every application
1:07:39you build in every possible way. There
1:07:41is a lot of variation here. Many things
1:07:43from software testing apply here. The
1:07:45good thing is that we have kind of
1:07:47covered this in our course. I have been
1:07:49conducting this course for the last
1:07:51four months. I have currently kept its
1:07:53interview weightage score at eight.
1:07:54Because I feel it is decently
1:07:56significant. At least one question is
1:07:58asked of you in the interview, and you
1:08:00are judged heavily based on this. And
1:08:02that question is mostly the same: if
1:08:04you built a RAG chatbot or an agent,
1:08:06how would you evaluate it? How do you
1:08:08know it is working correctly? And if
1:08:10you don't answer this correctly, people
1:08:12understand that you are basically "
1:08:14vibe-coding" your LLM-based
1:08:15applications. You don't have deep
1:08:17knowledge, and its on-the-job weightage
1:08:20is 10. Because without this, you cannot
1:08:23ship any application. What do you need
1:08:25to study? I won't discuss this much.
1:08:27Because literally, we have been doing
1:08:29this for the past four months. You need
1:08:31to cover it in three parts. You need to
1:08:33study LLM evals. How to evaluate LLMs.
1:08:36You need to study RAG evals. How to
1:08:38evaluate RAG applications. You need to
1:08:41study agent evals. How to evaluate
1:08:43agents. And you will find all three
1:08:45things in this course. LLM evals, this
1:08:48one—everything is in this course. It
1:08:51even has agent evals. Okay? So that was
1:08:54your module seven. Now let's talk about
1:08:57module eight. Module eight is also
1:08:59super important. And it's not that
1:09:01important right now. In the next two
1:09:02years, it will become even more
1:09:03important. I have a prediction here as
1:09:05well. Just as the field of cyber
1:09:07security emerged from software. As
1:09:10people started using software in their
1:09:12day-to-day lives, many attackers
1:09:14emerged to break them. So, a need for
1:09:17cyber security experts started to arise
1:09:19. Similarly, the more we use LLM-based
1:09:22applications, the more incidents of
1:09:24breaching them will increase, and the
1:09:26need for AI security experts will grow.
1:09:29So this job role will also emerge in
1:09:31the next three to four years, and
1:09:32specialists will be required here too.
1:09:34Currently, the same person, the AI
1:09:36engineer, is doing this work. The AI
1:09:38engineer is building the software. They
1:09:40are building the LLM-based applications
1:09:42. They are also evaluating and securing
1:09:44them. But my prediction is that in the
1:09:46next three to five years, these roles
1:09:48will diverge, and you will need experts
1:09:50in all three. But yes, for now, you
1:09:53have to learn everything. So AI
1:09:55security is super important. Interview
1:09:57weightage is currently a bit low.
1:09:59Because someone might ask one or two
1:10:01questions like, how will you secure it?
1:10:03What kind of threats are in your
1:10:05application? There isn't too much
1:10:07grilling here, and for on-the-job
1:10:08weightage, I'm currently giving it a
1:10:10seven because it's not being discussed
1:10:12much, but this score will increase in
1:10:14the next year or two. So here, you
1:10:17first need to understand what the
1:10:18threats are if you are building an
1:10:20LLM-based application. From which
1:10:22angles it can be attacked. This is
1:10:24called AI threat modeling. It is a kind
1:10:26of framework. You draw a diagram of
1:10:29your agent or your RAG chatbot and
1:10:31place all its components there, and
1:10:33then you think about where attacks are
1:10:36possible. So this is called threat
1:10:39modeling. After that, you have to read
1:10:40about the different types of attacks.
1:10:42Prompt injection, data exfiltration,
1:10:45jailbreaking, and you need to study how
1:10:47to secure agents separately. And you
1:10:50have to study RAGs in a different way.
1:10:52Data privacy, PII leakage, how you
1:10:54attack your own agent system or RAG
1:10:56system to test for vulnerabilities—
1:10:59which is called red teaming—and AI
1:11:01governance, meaning how to modify our
1:11:04applications based on the rules in
1:11:06different regions to launch them in any
1:11:09country; you need to study all this,
1:11:11and for that, we have a course called '
1:11:14LLM Guardrails'. The good news is that
1:11:17we have already started working on this
1:11:19and have added it to your dashboard.
1:11:22Here it is, 'AI Security and Guardrails
1:11:24'. Currently, we have added three
1:11:26modules to it. There will be a total of
1:11:28six modules. So, we are expecting that
1:11:31by October 31st, which is like 14-15
1:11:33days away, we will complete this course
1:11:36. This is a very important course, and
1:11:38it can give you an edge over others
1:11:40because not many people are studying
1:11:42this right now. If you know it, you
1:11:45will have an advantage. Alright? So,
1:11:48let's do a quick recap at this point.
1:11:51We covered software engineering
1:11:52fundamentals. We learned LLM
1:11:53fundamentals. After that, we learned
1:11:55how to work with LLM APIs. We learned
1:11:58prompt engineering, RAG, AI agents,
1:12:00context engineering, and how to
1:12:03evaluate LLMs. We studied AI security.
1:12:06These are the things we have covered so
1:12:08far. By this point, you have learned to
1:12:10build different types of applications
1:12:12and also learned to evaluate them. You
1:12:15also learned how to secure them.
1:12:16Alright? Now comes the next level,
1:12:19where you need to take that application
1:12:21and deploy it. But before deployment,
1:12:24you need to know a lot of things. And
1:12:28that is where our next level comes in,
1:12:30which is AI System Design, which I feel
1:12:32is important, especially if you are
1:12:34someone at a senior level with
1:12:35three-plus years of experience; for
1:12:37them, this becomes super critical. And
1:12:40it is being asked a lot in interviews
1:12:42right now. So, AI system design becomes
1:12:44important because these are the people
1:12:47who decide how a system will be created
1:12:50, deployed, and managed. Alright? So,
1:12:52AI system design becomes super
1:12:54important. And here, you understand all
1:12:57the things you have studied so far from
1:12:59a more theoretical side and from
1:13:01different angles. Alright? You
1:13:03understand different considerations. So
1:13:06, here you will study about context
1:13:08again, but you will try to understand
1:13:10context from every angle. Model routing
1:13:13and gateways is a very important topic.
1:13:15Instead of using a single LLM API, you
1:13:18might end up using more than one.
1:13:21Because if you have a simple task, you
1:13:23will use a smaller LLM for it. If a
1:13:26complex task comes up, you will use a
1:13:27complex LLM for that. Using the same
1:13:29LLM for everything can increase your
1:13:31costs. It will cost you more. So, how
1:13:33do you use model routing? How do you
1:13:35implement LLM gateways? How do you
1:13:38optimize latency and inference? That is
1:13:40a big question. How do you cache things
1:13:43? So that you don't pay for the same
1:13:45things over and over. Caching can be
1:13:47done at different levels. So, how do
1:13:49you do that? How do you implement
1:13:51guardrails in your system? How do you
1:13:53build reliable LLM-based applications?
1:13:55You have to learn all of that.
1:13:58Basically, at this point, you need to
1:14:01learn how to plan a complete end-to-end
1:14:04LLM-based application. Okay? And
1:14:08unfortunately, at this moment, we don't
1:14:10have a course at CampusX to teach you
1:14:12this. But our goal is to teach you this
1:14:16course within the next four months. But
1:14:20still, if you want to study from day
1:14:22one, I would recommend one resource.
1:14:26This is the only external resource we
1:14:28are including here. Because I have read
1:14:31this particular book myself. In fact,
1:14:33it is kept right here behind me. The
1:14:36name of this book is "AI Engineering:
1:14:38Building Applications with Foundation
1:14:39Models" by Chip Huyen, a very, very
1:14:41good author. She has another book, "
1:14:44Designing Machine Learning Systems,"
1:14:46which is basically about ML system
1:14:47design. Her content on system design is
1:14:50always very solid. So, if you want to
1:14:52learn AI system design today, this is
1:14:54the best resource you have. We will
1:14:57also create a course, and hopefully, we
1:15:00will create a great course. But at this
1:15:03point, if you want to learn, you should
1:15:04read all the chapters of this book. In
1:15:06fact, here it is, I have the book right
1:15:08here. This is the book. AI Engineering.
1:15:11A very good book. So, I am not
1:15:13discussing this syllabus in detail
1:15:14right now because this course isn't
1:15:16built yet. So, this syllabus isn't
1:15:17final. It is possible that there might
1:15:19be some minor changes. But if you want
1:15:21to start today, this is a very good
1:15:23book to start with. Okay? This is AI
1:15:25system design, level number nine. Level
1:15:28number 10 is also something that is
1:15:29very important. Oh, by the way, I
1:15:31haven't mentioned this. I have given
1:15:32this an interview weighted score of
1:15:34nine out of 10. Because this is a
1:15:35dedicated round. System design round.
1:15:38You are grilled a lot about this. So,
1:15:39you must know it. And the on-the-job
1:15:41weightage is also eight out of 10.
1:15:42Especially if you are someone going for
1:15:45a senior level role, you have to plan
1:15:47the system before anything starts. So
1:15:49it is kind of important. At slightly
1:15:50junior levels, it is not as important.
1:15:52That is why I have set the score to
1:15:54eight. Next is LLM Ops. Why do you have
1:15:57to study both? Let me explain it to you
1:15:59. I am asking you to study AI system
1:16:00design as well. I am also asking you to
1:16:02study LLM Ops. The job of both is the
1:16:03same. Building and deploying a system
1:16:07in a proper way. Running it reliably.
1:16:09The job of both is the same. But I am
1:16:11telling you to study both because both
1:16:13are important. Again, the example for
1:16:16this is, let's say you are given the
1:16:19task of building and running an airport
1:16:21. So there will be two stages in this.
1:16:23Right? You will first do a detailed
1:16:26analysis. A detailed study to
1:16:29understand and plan how you will build
1:16:31an airport. You will go and study where
1:16:34to acquire the land. You will think
1:16:37about where on that land I will keep
1:16:39space for passengers. Where will I
1:16:41build the runway? Where will my air
1:16:43traffic controller be? You will plan
1:16:45all of this. Then you will plan how
1:16:48many flights I expect daily, how many
1:16:50flights will leave, how many people
1:16:51might come, and how many security
1:16:53personnel are needed. So this entire
1:16:56planning that is happening is the same
1:16:58as what you do in AI system design.
1:17:01Before starting anything, you plan
1:17:03everything, like I will set up an LLM
1:17:06gateway where I will redirect. I will
1:17:10do routing so that this request should
1:17:11go to this model. This should go to
1:17:13this model. I will batch the requests
1:17:16or I will implement this kind of UI.
1:17:19All this planning you do before writing
1:17:21code is AI system design. After that,
1:17:24the second stage comes where you
1:17:25actually implement that plan. You start
1:17:28building the airport and run it on a
1:17:30daily basis, that today these flights
1:17:31will land. These flights will take off.
1:17:35These many passengers are arriving. We
1:17:37are doing their security testing here.
1:17:39This day-to-day running and operations
1:17:41is what LLM Ops is. That is why it is
1:17:44said that whoever studies both will be
1:17:46able to operate in the best way. But if
1:17:48you only have to study one of these two
1:17:51, then even if you only study LLM Ops,
1:17:53you will be able to do your job. But
1:17:55someone who has studied AI system
1:17:57design will work in a better way.
1:17:59That's the only difference. That's why
1:18:01I am asking you to study both. Because
1:18:02this is our comprehensive roadmap. So
1:18:05yeah, the interview weightage is seven.
1:18:07Not too many questions are asked
1:18:09because this is something that matters
1:18:11more on the job. That is why the
1:18:13on-the-job score is nine. And again, we
1:18:15won't discuss much here either. Because
1:18:17literally, you have been studying this
1:18:19course for the last three to four
1:18:20months. Himanshu has been teaching you.
1:18:21So all the stages that were taught to
1:18:24you—prompt management, evaluation,
1:18:27experiment tracking, security and
1:18:30guardrails, dockerization, CI/CD,
1:18:33post-deployment—you have to study all
1:18:36those things here. This is exactly what
1:18:38is mentioned here. And this course is
1:18:40now complete for you. Complete, meaning
1:18:42a few things might be left, I guess. So
1:18:44you will find those here. I guess it is
1:18:46still within the Insiders course. We
1:18:48haven't published it separately. I
1:18:50guess whatever it is. Fine, you guys
1:18:52know. You are already using it anyway.
1:18:54So yeah, this is level number 10. After
1:18:56that comes level number 11. Up to this
1:18:59point, we have covered the entire
1:19:01spectrum. We have learned to build
1:19:03applications, evaluate them, and secure
1:19:05them. We have also learned to deploy
1:19:07them end-to-end. We have also learned
1:19:09to manage them. Now comes the 11th
1:19:11module. These are all optional modules
1:19:13now. 11, 12, 13 are kind of optional
1:19:15modules. The 11th is Managed AI
1:19:18Platforms. Here, basically, you need to
1:19:22learn about things like AWS Bedrock,
1:19:24Azure Foundry, or Google Vertex AI.
1:19:28Basically, these cloud providers have
1:19:30given you an end-to-end platform. Where
1:19:33you can do all of these things.
1:19:34Everything that you have studied so far
1:19:36, how you will do all that on the AWS
1:19:38platform—that is what this module is.
1:19:41So if you are interviewing for a
1:19:43company where their entire business
1:19:46runs on an AWS stack, then this module
1:19:48becomes relevant. If you are working
1:19:51for a startup that is building its own
1:19:53entire tech stack, then this module is
1:19:55not important. That is why its
1:19:57interview weightage score is low, and
1:19:59its on-the-job weightage is also low.
1:20:00Okay? And I won't describe this much
1:20:02right now either. We haven't covered
1:20:04this yet, and it is also a bit lower on
1:20:07our priority list. So, we will cover
1:20:09this later. Most likely, this will be
1:20:11covered after December. At least we
1:20:13will cover one of the three platforms
1:20:15thoroughly. We will take a poll on
1:20:17whether to do Bedrock or Azure Foundry.
1:20:19But we will cover one in detail. But
1:20:22yeah, again, this is something that is
1:20:24very specialized for certain companies,
1:20:25not useful for all companies. So, in
1:20:27that sense, it is somewhat skippable.
1:20:29Okay? So this is level number 11. Level
1:20:32number 12 is also not very important.
1:20:35It is multimodal, one beyond text.
1:20:39Everything we were studying until now
1:20:40was mostly around LLMs and textual data
1:20:42. Building chatbots, agents, RAG chat,
1:20:46everything was around textual data. But
1:20:49it is possible that you might work in a
1:20:51company whose use case is beyond text.
1:20:53Now, here too, you might have to study
1:20:56a little. So here you will learn about
1:20:58vision models or vision capabilities.
1:21:01Of LLMs. You will learn about audio
1:21:03capabilities. In vision, video is also
1:21:08included. And besides that, I have
1:21:11added one or two minor things here,
1:21:13like how to build long-running tasks or
1:21:15long-running scenarios, which is
1:21:17mentioned here. And Edge AI, basically
1:21:20AI that runs on your mobile devices or
1:21:23edge devices. How to build those
1:21:25systems? Again, a very specialized job
1:21:27role. Very few companies are working in
1:21:29this domain. So, that is also kind of
1:21:31mentioned here. And again, we don't
1:21:33have any course related to this right
1:21:35now. But hopefully, we can build it in
1:21:38the future, in the next six months.
1:21:40This is a bit lower on our priority
1:21:42list, and its interview weightage is
1:21:44also low. Because if a company doesn't
1:21:46work in vision or audio, they won't
1:21:47even ask you questions about it. They
1:21:49will only ask you questions around
1:21:51textual data. And the on-the-job
1:21:52weightage is also low. Because for the
1:21:54same reason, if a company is not
1:21:56working on it, they won't give you such
1:21:57projects. But again, for completeness,
1:21:59what did I tell you? This is the most
1:22:01comprehensive roadmap that I have
1:22:03created. So for the purpose of
1:22:05completeness, I have mentioned it here.
1:22:07Okay? After that comes the 13th. 13th
1:22:10is fine-tuning. Which is talked about a
1:22:14lot. Let's discuss this in a bit more
1:22:16detail. So, fine-tuning is something
1:22:18that is important from an interview
1:22:20point of view. Because I have
1:22:22personally seen that students are asked
1:22:24about fine-tuning techniques in
1:22:26interviews. So, the interview weightage
1:22:29score is seven, but the on-the-job
1:22:31weightage score is only three, because
1:22:33based on everyone I've talked to, there
1:22:35are very few AI engineers who have done
1:22:37fine-tuning or have needed to do it.
1:22:39Most of the time, you can achieve the
1:22:42same thing with the help of intelligent
1:22:44prompting and RAG. It is very rare that
1:22:47you actually need to do fine-tuning.
1:22:50Okay? So, there are fewer companies
1:22:51that require this work. But again, it
1:22:53is important for interviews, so you
1:22:54will have to study it. Here, again, are
1:22:58the same famous topics: supervised
1:23:01fine-tuning, LoRA, QLoRA, and some
1:23:03study around alignment. Distillation is
1:23:08something that is being talked about a
1:23:10lot these days, so distillation, how to
1:23:13evaluate the fine-tuned models you've
1:23:15built, and lastly, how you serve the
1:23:17adapters you get after fine-tuning in
1:23:19production scenarios. Mostly, this is
1:23:23what you need to study in the
1:23:25fine-tuning module, and again, this is
1:23:27something you won't find on our CPS X
1:23:29website. Because we haven't created any
1:23:32resources around this. But this is
1:23:34again something that is very high on my
1:23:36personal priority list. Out of
1:23:37everything we have discussed so far, I
1:23:39am going to teach three things. One is
1:23:42this LLM 1001, one is AI System Design,
1:23:44and one is this fine-tuning. I will
1:23:47teach these three things within the
1:23:48next six months. So, I have a two-month
1:23:50distribution in mind, and the order is
1:23:52roughly this as well. LLM 1001, then AI
1:23:54System Design, and then fine-tuning. I
1:23:57hope to cover this by the end of
1:23:59February or March. Okay? So, I want to
1:24:02give you a bit of clarity because many
1:24:04people ask about this. So, I feel like
1:24:06I should clarify what is going on in my
1:24:09mind. So yeah, this was 13, and now the
1:24:11last two modules. These are nothing
1:24:13different. These are for your
1:24:15preparation. These are for portfolio
1:24:16building. So, 14th is Project. The
1:24:19interview weightage score is 10 out of
1:24:2010. Because in interviews, you are
1:24:23literally grilled about your project.
1:24:26The better you can explain it, the more
1:24:28solid your chances are of cracking the
1:24:29interview. The on-the-job weightage is
1:24:31not as high. Because the projects you
1:24:34build as students and the projects you
1:24:35have to work on in the industry are
1:24:37kind of different. So you can use that
1:24:40experience a little bit. But it doesn't
1:24:42have much utility. So here, I have made
1:24:46a list of about five or six good
1:24:48projects. You can go and take a look.
1:24:53Currently, we haven't added much
1:24:54content here. My plan is to build these
1:24:56projects and add them to our platform.
1:24:59So from December onwards, we will build
1:25:02these projects and add them to the
1:25:04platform. Right now, we don't have
1:25:06these projects. But yeah, this is like
1:25:08the plan to do within the next 6 months
1:25:10. And lastly, 15th is interview
1:25:13preparation, based on whatever you have
1:25:16studied so far, we have divided the
1:25:18interview preparation well here. There
1:25:20is a separate system section for RAG.
1:25:22There is a separate one for Agent AI.
1:25:24There is a separate one for Fine-tuning
1:25:25. There is a separate one for LLMOps
1:25:26and production debugging. There is a
1:25:28separate one for AI system design. How
1:25:29to handle current things that are
1:25:32happening right now. There is a
1:25:34separate section for that. There is a
1:25:35separate one for project and portfolio
1:25:37defense and a separate one for
1:25:38behavioral. And for this, we have a
1:25:40course. We even conducted a course for
1:25:42you. Varun took that course. Who is
1:25:44working as an AI Engineer. So I guess
1:25:47you will find that course. This AI
1:25:50Engineer Interview Preparation.
1:25:53Although we haven't taught interview
1:25:54questions here, we have explained
1:25:56everything very well, like what you can
1:25:58expect in an AI engineering interview.
1:26:01What kind of questions can come. How to
1:26:03answer those questions, how not to
1:26:04answer them. So we had taken six or
1:26:07seven sessions on this. So if you study
1:26:09that, it will kind of help you with
1:26:12this part. Okay? So in a nutshell, this
1:26:14is the comprehensive road map of AI
1:26:17engineering. And as I told you in the
1:26:19beginning, this is like the super set.
1:26:21There is nothing outside of this now.
1:26:24Having worked in this domain for two
1:26:26years, I have realized that if you
1:26:28learn all these things, you will become
1:26:30an absolute expert in this entire
1:26:32domain. You don't need to study
1:26:34anything outside of this. The only
1:26:36downside is that this is too much. If
1:26:40someone goes to study all this, it will
1:26:41take them a year, and the problem with
1:26:43that is that something new will come up
1:26:44in that one year. So in a way,
1:26:46technically, it is not possible to
1:26:48complete everything that is provided
1:26:49here in this road map. So why did I
1:26:52teach you this entire thing? Spending
1:26:53an hour and a half. Mostly because I
1:26:56wanted to give you an idea of what the
1:26:57boundary is within which you need to
1:26:59operate. Okay? So, this is your full
1:27:02roadmap. Do not pursue this. Now, I
1:27:05will show you a subset of this that I
1:27:07think will efficiently cover 70-80%of
1:27:10this entire roadmap in half the time.
1:27:13Okay? So, I want to show you that, and
1:27:16it is our accelerated roadmap, which we
1:27:18are calling 100 Days of AI Engineering.
1:27:21And I feel that most of you will do
1:27:23this version rather than going for the
1:27:26first one. So look, what I have done
1:27:30here is applied some thought, and I
1:27:32have utilized my teaching experience,
1:27:35which is now perhaps 15 years old, to
1:27:38compress this entire thing by half
1:27:40while retaining 60, 70-80, maybe 90%of
1:27:43the value that was in the original
1:27:46roadmap. Okay? So now, where there were
1:27:4916 levels there, here there are only
1:27:52six plus two, eight levels. So it has
1:27:54literally come down to 50%. And I will
1:27:56give you a rough idea of what is there,
1:27:58and then we will discuss its specifics.
1:28:00So, LLM fundamentals are still there.
1:28:02Uh, I will actually open this in a new
1:28:04tab. Just to show you what the
1:28:06comparison is. So, I have removed this
1:28:11level. Level zero, which was this level
1:28:14. I have removed this because I am
1:28:17assuming that whoever studies the
1:28:19compressed version will come prepared
1:28:22with this. Okay? This is my assumption.
1:28:25So that is why we are starting directly
1:28:28with LLM Fundamentals, which is this
1:28:30one, level number one. Now, in this too
1:28:32, what have I done? I have kept very
1:28:34few things. Just enough to get you
1:28:37started on how LLMs work. Tokens,
1:28:40context, and latency. Three important
1:28:42things you should know about. I have
1:28:45included a little discussion about
1:28:46multimodal here, because if needed, it
1:28:48is kind of skippable. If you do not
1:28:50want to do it, you can skip it too. How
1:28:52to control output using temperature or
1:28:54reasoning models, that has been
1:28:56compressed into one module. Where LLMs
1:28:59can fail and how to choose models. That
1:29:02is all you have to study. And here, I
1:29:04have also mentioned what things I have
1:29:06removed and why I have removed them,
1:29:08and where you will find them in the
1:29:10original roadmap. For instance, I have
1:29:12removed Attention and Transformer
1:29:14Internals. It was in the original
1:29:16roadmap and is interesting, but it
1:29:17doesn't change what we are building or
1:29:19the changes we are making. But if you
1:29:21want to study it, you can always go to
1:29:23this place and study there. Okay? So, I
1:29:25have drawn those connections. So yes,
1:29:27LLM fundamentals still exist because I
1:29:29think they are important, but in a
1:29:31compressed way. Second comes this part,
1:29:34which was working with LLM APIs; it’s
1:29:37the same, but I’ve made some
1:29:39reductions here as well. Although I
1:29:42couldn't remove much because most
1:29:43things aren't redundant. That’s why
1:29:46you will see that 80%of the content
1:29:48still exists. What has been removed is
1:29:51mentioned here. I have removed calling
1:29:53the API providers directly. I have
1:29:55taught how to do it directly through
1:29:56LangChain. I haven't included retry
1:29:58logic and such things yet. Because
1:30:01those are useful when writing high-end
1:30:03software. So, it is kind of skippable.
1:30:05I have removed reading about LlamaIndex
1:30:07and things like that. And I have
1:30:08removed the comparison of orchestration
1:30:11frameworks. Okay? So, I kept the things
1:30:13that are very important. After that,
1:30:15third comes prompt engineering. Some
1:30:17things have been kept in this as well.
1:30:19Zero-shot, few-shot, system prompts,
1:30:22and so many things have been removed.
1:30:24Okay? So, prompt engineering has also
1:30:26been trimmed down a bit. Now comes the
1:30:28main thing where I applied my focus.
1:30:31You have to study RAG. But this time,
1:30:33I’ve asked you to study RAG in a
1:30:34slightly different way. Now, we have
1:30:37made it so that in the 100 Days of AI
1:30:39Engineering compact roadmap, the
1:30:42accelerated roadmap, you will learn
1:30:44everything through one application. So
1:30:47here, first you will build a RAG
1:30:49application, and then you will convert
1:30:52that same RAG application into an
1:30:54advanced RAG. Then you will evaluate it
1:30:56right there. And you will secure it
1:30:58right there as well. So, what happened
1:31:00is that you are studying RAG in four
1:31:02sections. First, you built a simple RAG
1:31:05. Then you added advanced capabilities
1:31:07to it. Then right there, you learned
1:31:09how to evaluate it. Which basically
1:31:11means that in the compressed roadmap,
1:31:13evaluation is not a separate module.
1:31:15You are studying it along with RAG. And
1:31:17similarly, once you have evaluated it,
1:31:20securing it is also something you have
1:31:22to study right here. So, AI security is
1:31:24not a separate module either. So
1:31:26basically, I broke down those two
1:31:28modules and put them here into RAG.
1:31:30Okay? So, you built a RAG application,
1:31:32made it simple, then advanced,
1:31:34evaluated it, and secured it, giving
1:31:37you a complete picture of RAG. This
1:31:40will take some time; I've literally
1:31:42allocated 30 days out of 100 here, so
1:31:44you need to spend time on this. Then
1:31:46comes the fifth module, which is Agents
1:31:49, and the idea here is that we’ll
1:31:51turn that same RAG application into an
1:31:54agent. Basically, we will build an
1:31:55agentic RAG application. So, it will
1:31:57have both RAG capabilities and agentic
1:32:00capabilities. Meaning you will
1:32:01literally keep working on the same
1:32:03application. Now, what do you need to
1:32:05do here? You have to build an agent.
1:32:07You then need to add advanced features
1:32:09to it. Such as MCP tools, or
1:32:11incorporating the concept of memory
1:32:12right there. You have to study a bit of
1:32:14context engineering through it as well.
1:32:16If you want to implement a multi-agent
1:32:18architecture, you are doing that within
1:32:20it too. And once you have done all of
1:32:21this, you are also learning how to
1:32:23evaluate it right there. So, we learned
1:32:25agent evaluation here, and we also
1:32:27learned how to secure it right there.
1:32:30The biggest compaction has come from
1:32:33the fact that the two modules you were
1:32:35supposed to study—one was evaluation,
1:32:37LLM evaluation, and the other was AI
1:32:39security—You don't need to study
1:32:41these two separately. You will
1:32:44literally learn both of these things in
1:32:46the process of building the application
1:32:48, studying RAG, and studying agents. So
1:32:51, you have to study two fewer modules.
1:32:53That’s the breakdown. Right? So, what
1:32:55have you learned so far? LLM
1:32:56fundamentals, working with LLM APIs,
1:32:59prompt engineering, RAG, including
1:33:02simple RAG, advanced RAG, evaluation,
1:33:05and security. In agents: basic agents,
1:33:08advanced agents, where you studied
1:33:09memory, context, and all that. Then
1:33:12evaluating them, and then securing them
1:33:14. So, you have literally covered
1:33:15everything up to this point. You have
1:33:18covered up to module eight here in just
1:33:21five modules. Right? After that comes
1:33:25the sixth thing, which will give you a
1:33:27taste of both AI system design and
1:33:30LLMOps. I have named it "Shipping to
1:33:33Production." Basically, the app you
1:33:35have built so far—the agentic RAG
1:33:37application that you have already
1:33:39created, evaluated, and secured—Now
1:33:41you need to put it into production. So,
1:33:44I merged both concepts in one place.
1:33:46Both AI system design and LLMOps. So it
1:33:49starts with AI system design. We have a
1:33:51brief discussion on how we will take it
1:33:53to production, and after that, the
1:33:55actual LLMOps work begins. We build the
1:33:57front end. We do packaging and
1:33:59deployment. We put it into production
1:34:02and monitor it. We monitor cost, speed,
1:34:05and scale. And we improve that
1:34:07deployment a little further. We
1:34:08implement model gateways or replace
1:34:11them with open models. We have a little
1:34:14discussion about fine-tuning right here
1:34:16, to see if fine-tuning the model we
1:34:18used would have given us better
1:34:19responses. So you are covering this
1:34:22entire thing in a single go. You are
1:34:24spending time here. You are spending
1:34:26around 15 days. But the goal is to take
1:34:28the application you have built so far
1:34:30to production and have all these
1:34:32discussions during that production
1:34:34phase. So literally, both of these blue
1:34:36ones have merged. In that sense, it has
1:34:38become a bit compressed, and once you
1:34:40have put this application into
1:34:41production, your job is done, and from
1:34:43here, you only need to do two more
1:34:45things. First, leave the running
1:34:48project you just built, and now build
1:34:51two or three more projects from the
1:34:54given ones using the same flow. If it
1:34:57is a RAG application, take the RAG
1:34:59application flow. If it is an agentic
1:35:01application, take the agentic
1:35:02application flow, and build two or
1:35:04three projects from these using the
1:35:05same flow. Okay? And these are
1:35:07excellent projects. Everything is
1:35:09explained in detail. What to build,
1:35:11what to show, which skills, how to
1:35:12present it—everything is written here
1:35:14. You can select two or three, and it
1:35:16is more than sufficient. And once you
1:35:18have built the projects. After that,
1:35:20the interview section has also been
1:35:21simplified. I have literally searched
1:35:23for 100 good questions, divided them
1:35:25into different sections, and written
1:35:27them down here. These are the most
1:35:29frequently asked questions today. These
1:35:31are the 100 questions. If you prepare
1:35:33these 100 questions. I feel that you
1:35:35will be able to answer seven out of ten
1:35:37questions in any interview. So this is
1:35:40the compressed 100-day format that I
1:35:42have extracted from that original
1:35:45roadmap. So if we cover the original
1:35:49roadmap sometime in the future, uh, in
1:35:52the next 6 months, then maybe I'll take
1:35:55a 100-day sprint and literally teach
1:35:57you this compressed curriculum again
1:36:00here or on YouTube in those 100 days.
1:36:04That's the goal, but again, it's a long
1:36:05way off, so I'm not making any promises
1:36:07right now. But yeah, this one is a much
1:36:09better flow. If you want to cover
1:36:12things in a shorter span, this is a
1:36:14much better flow. Okay, start with LLM
1:36:18fundamentals, learn to use LLM APIs,
1:36:21learn a bit of prompt engineering, and
1:36:25learn evaluation and security while
1:36:28you're studying RAG. Learn evaluation
1:36:32and security while building agents, and
1:36:34learn how to put them into production
1:36:35properly. This flow can give you 80%of
1:36:39the knowledge we covered in detail in
1:36:42the comprehensive roadmap. These are
1:36:45the two roadmaps. And one last thing
1:36:48before we end the session. Why you
1:36:50should follow this particular roadmap
1:36:53that I gave you today. Uh, many other
1:36:56roadmaps exist. Our promise is that we
1:36:59will keep updating this roadmap
1:37:01frequently. Because I have been doing
1:37:03this for the last three or four months.
1:37:05And this roadmap is literally the North
1:37:07Star for our entire company. We are
1:37:09literally all working together on this
1:37:12very roadmap. Working on different
1:37:14sections of this same roadmap. As I
1:37:15mentioned, I will be taking three
1:37:17courses. Himanshu is assigned a couple
1:37:19of courses. Rupender is assigned
1:37:21courses. Ajay is assigned courses. We
1:37:23are hiring some more teachers. We will
1:37:25assign courses to them as well. Uh,
1:37:27some people will handle projects. Some
1:37:29people will handle interviews. So the
1:37:30idea is to cover this roadmap
1:37:32end-to-end so that within six months,
1:37:34we can provide you with the complete,
1:37:36comprehensive roadmap. And hopefully,
1:37:39conduct a live course for the shorter
1:37:41version as well. So, that's the goal.
1:37:44And the second reason to follow this
1:37:46roadmap is that we are giving you both
1:37:48versions. Most people will give you the
1:37:50shorthand version. I have given you a
1:37:52complete overview of what the superset
1:37:54is. And then I have also given you the
1:37:56shorter one. And third, if you are
1:37:58already in the CampusX ecosystem.
1:38:01Meaning you have already taken our
1:38:03courses or prefer our teaching style,
1:38:04then you should follow this roadmap.
1:38:08Because we are using this roadmap to
1:38:10plan all of our future execution. So
1:38:13these are the reasons why you should
1:38:15follow our roadmap. There is no other
1:38:18problem. Whichever roadmap you like,
1:38:20whichever clicks for you. The end goal
1:38:22is to get a job. It is a difficult
1:38:24situation, a difficult market. Cracking
1:38:26a job is the highest priority. It
1:38:28doesn't matter which roadmap you come
1:38:29from. I am not here to sell anything.
1:38:32In fact, if you find a good roadmap,
1:38:34you tell us too. We will pick good
1:38:36things from them and add them to our
1:38:38roadmap. But yeah, our goal is to
1:38:41provide you with the best, and that is
1:38:43what we have been trying to do for the
1:38:45last four months with the Inside
1:38:47Assessors program. So yeah guys, this
1:38:51link has reached you here on the
1:38:53website. You can go, you can follow it,
1:38:58and if you find some good free
1:39:01resources, send them to us. We will
1:39:05create a separate section here and add
1:39:08them. Currently, we are only showing
1:39:12the campus ones here. But we want to
1:39:15add free resources here too. Because it
1:39:17feels a bit odd to only send our own
1:39:19course. So if you know of any good
1:39:22resources that you have already used,
1:39:24send them to us. We will test them at
1:39:27our level and we will add that. Okay?
1:39:29So that we don't get blamed for just
1:39:30selling our own course. So yeah guys,
1:39:32that's it. This is what I had to cover
1:39:34today. Tomorrow will be an important
1:39:36session too. I want to discuss the
1:39:39report that I showed you with you. And
1:39:42I want to discuss your career-related
1:39:45doubts in a bit more detail. So try to
1:39:47attend tomorrow's session. It would be
1:39:50nice.

About this transcript

This page contains the full transcript of AI Engineering Full Roadmap vs 100-Day Sprint: In-Demand Skills, Job Market Data & Curriculum by CampusX, generated from the public captions YouTube serves with the video. The transcript has 17,926 words across 2,636 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.