FAANG PM Reveals How to Build AI Agents (and Get Paid $750K+)

1. 0:00 – 1:38

   Intro

   1. AGAakash Gupta0:00

      The future of product management and building products is building AI agents, but most people are gatekeeping the knowledge

   2. MYMahesh Yadav0:07

      In past, a developer need to read 20 API documents, but now all you are doing is-

   3. AGAakash Gupta0:12

      And then after that, you worked at so many of the other FAANG places

   4. MYMahesh Yadav0:16

      That's the old world. The new world is-

   5. AGAakash Gupta0:17

      Mahesh Yadav has been a product manager at Meta, Amazon, Microsoft, and Google. In today's episode, he's gonna break down exactly how to build AI agents

   6. MYMahesh Yadav0:26

      So Microsoft is a dreamland for people who want to go build without caring how it will make business. That's the opposite of that, Amazon.

   7. AGAakash Gupta0:34

      How did you get so good at the PM interviews, though?

   8. MYMahesh Yadav0:36

      These three skills were essential for being a successful PM

   9. AGAakash Gupta0:39

      The future of product development is agentic. You need to be able to build AI agents into your products. The AI agent PM role is one of the fastest-growing and highest-paying roles out there. These roles pay $750,000 plus. Today we're breaking down the exact 18-month roadmap you need

   10. MYMahesh Yadav0:58

       Nobody can stop you. Next iteration, you are even doing better than their PMs. Google will hire you, right?

   11. AGAakash Gupta1:02

       What do FAANG companies like Google look for in these interviews?

   12. MYMahesh Yadav1:06

       All of us should build a reviewer. People who are new in AI lose lot of their brand in the beginning

   13. AGAakash Gupta1:11

       A lot of people, they wanna break into these companies. They want to get this 300,000-plus PM AI agents job. So let's start with your own story. What is the salary and compensation for these AI agent PM roles? Mahesh Yadav is one of the most seasoned product leaders when it comes to AI agents, and so this episode is a masterclass. Without further ado, let's get right into it. Mahesh, what are we gonna cover today?

2. 1:38 – 2:37

   What Makes an AI Agent PM

   1. MYMahesh Yadav1:38

      Today, I'm gonna show you how you build agents, not only a fancy front end in Loveable, but also a back end that you can scale with multi-agents. Also, I will talk about how to get 300K-plus jobs in AI, and especially agentic AI, and share any secrets that I have learned throughout the years.

   2. AGAakash Gupta1:57

      Amazing. What does it take to become an AI agent PM?

   3. MYMahesh Yadav2:00

      Oh, [laughs] I know it, uh, sounds hard, but, uh, it's three simple things. One, we are looking for people who have built in AI. Second, we are looking for people who have done AI PM transactions, which includes, like, how you dealt with data, how you dealt with models, evaluations, and scaling by iteration. And last is just generic skills around scale, ambiguity, and experimentation, because this is new and these three are becoming essential or table stakes now

   4. AGAakash Gupta2:33

      So I don't want people to have to wait too long. Let's get right into how to build

3. 2:37 – 8:27

   Live Demo: Building Competitive Analysis Agent

   1. AGAakash Gupta2:37

      these AI agents, both back end and front end.

   2. MYMahesh Yadav2:40

      Sounds like a plan. So I will be using two major tools. Uh, we do this, uh, uh, uh, with new people, that we show, uh, them how can they build their AI agents. So for back end, I will be using this tool called Langflow. Langflow is a tool which allows you to build no-code, highly complex or simple agents. So let's start building our back end first, and I will be building a simple tool which you can use for competitive analysis. So as a PM or as anybody who's trying to build or analyze markets, you need to analyze competition. So we'll be building a tool which allows you to do competitive analysis

   3. AGAakash Gupta3:17

      So to go through all this content, let's get started with building an AI agent, shall we?

   4. MYMahesh Yadav3:21

      Yeah. That sounds amazing. So let me just tell you how we're gonna build it. We're gonna use Langflow, which is a no-code tool which you're looking at right now. It allows you to build highly complex or simple agents without writing a single line of code. And for the front end, we will be using v0. This is like Loveable or Bolt. You just write prompts, and it creates a awesome web app or front end for you. Then I will show you how can you combine a very powerful back end in Langflow to a front end in v0. So that's-

   5. AGAakash Gupta3:52

      Awesome

   6. MYMahesh Yadav3:52

      ... our plan today.

   7. AGAakash Gupta3:53

      Perfect. And I'm just curious, before we even dive into that, uh, how did you choose Langflow and v0?

   8. MYMahesh Yadav4:00

      Yeah, that's a good question, right? So I think when I was starting, there was ... I was looking for no-code tools because, uh, once you start writing code, you need to learn editor, you need to learn coding skills, you have to have... And suddenly the barrier to entry is so large that most of the people for- doesn't cross that barrier. That's why I was looking for tools which are easy. So one is ease. Second is capability. So Langflow was quite capable. It had integration with all model providers, so I can choose Azure, AWS, or Gemini Google models.

   9. AGAakash Gupta4:33

      Nice

   10. MYMahesh Yadav4:33

       And third is this ability to get access to the code if we want to go to production, and that also was there. Uh, I also look at n8n, and there was one cool thing was that it was completely free-

   11. AGAakash Gupta4:46

       Yeah

   12. MYMahesh Yadav4:46

       ... and n8n was charging, like, after two weeks, and I didn't have a free sub- subscription, so landed on Langflow. Okay, then let's get started. So the first thing first, what you do is you can go to Langflow.data. Uh, Langflow is this tool, and DataStax is the company which b- bought Langflow, and now IBM bought them. [laughs] So it's a journey. You will see this happen again and again in AI. So what we will be building is a competitive analysis tool. You can get to Langflow. The first time you go, you might need to create an account or log in with your Google credentials. Then you will land here. You can just click New Flow and choose Simple Agent. With this agent, you can actually build an end-to-end agent. So let me just get you started there. So this is a simple agent. You can use this, or let's use this to create our own agent. So this one takes a URL input, a calculator input, and we are building a competitive analysis tool.

   13. AGAakash Gupta5:44

       Yep

   14. MYMahesh Yadav5:44

       So what we need as input. And I think that's the first thing we all need to learn. If we can start talking in terms of input/output, and what process or what things we need to process, that would be a good product requirements or a interview, uh, or an good way to handle an interview is first up, anybody ask a question, see what are the inputs, what are the outputs.And I jot it down for you guys. So what you can see is if you're building your, uh, your agent, first step you need is you need which competitors you're gonna compare. So you need an input a-as a name of your competitors, what tools your agent will be using. It needs at least a search tool so that it can do all the search on internet and find that information that you're looking for. And then obviously you need agent to process and which requires you to have a prompt with special instructions, which I will show you a little bit tips and tricks on writing these prompts. And then what is the output? So input is name of competitors and output hopefully would be a table format, and tool is a search tool. Okay, great. If we need to build this, then let's go back to our Langflow, and our input is basically twofold. So this agent you can also drag and drop. So that was the cool thing I was talking about. So two inputs. So let's get those two inputs in. Text input, I will edit it. I will say Competitor one, Competitor A, and in the text input I can just say OpenAI. Great. And then I can delete this. I'm not looking for chat input, and I can just Control C, Control V, and I can just say Comp B, and let's compare that with Google 'cause that's why it was created, hopefully.

   15. AGAakash Gupta7:38

       [chuckles] So in this case, we're a hypothetical PM where?

   16. MYMahesh Yadav7:44

       You're a PM which is trying to do a competitive analysis across this company, and let's say you're trying to see which APIs to use or which to invest in, or you're looking at this company that you're starting a new model company, and you are trying to see who are out there-

   17. AGAakash Gupta8:00

       Okay

   18. MYMahesh Yadav8:01

       ... in creating these models.

   19. AGAakash Gupta8:02

       Yeah.

   20. MYMahesh Yadav8:02

       So maybe we need a new model which helps us because none of them are good enough anyways.

   21. AGAakash Gupta8:07

       [chuckles] Yep.

   22. MYMahesh Yadav8:10

       So then you need a prompt. So you can put a prompt one, and this is your prompt template, and now it requires you to copy-paste or get to your-- what message you want to write in your prompt. And prompt writil-

4. 8:27 – 12:03

   Writing System Prompts

   1. MYMahesh Yadav8:27

      writing is the art, I think is these, these days.

   2. AGAakash Gupta8:30

      Yeah.

   3. MYMahesh Yadav8:31

      So let me show you a good prompt.

   4. AGAakash Gupta8:33

      Yeah. The system prompts in sp- in particular for your AI agent are everything.

   5. MYMahesh Yadav8:38

      So you see the prompt. Let me just copy-paste the prompt because I'm not gonna type, but let me talk about what it does and how you can write a good prompt, a good prompt structure. The first thing you talk to all your agent is your role. So you're saying, hey, you're a professional business analyst with expertise in corporate research and competitive benchmarking. Now your instructions, compare the two companies, given company A, company B. These brackets are telling it that this information is coming from outside. I will show you in a second. Once you put it in bracket, the prompt structure will change, and it will open up two ports so that you can pass this information from outside. And taking help from Tavily. Oh, what is Tavily? I think it's a tool for doing AI research or search. I will attach that in our flow in a second, and then put the structured in a markdown format that includes the following attributes. And then you're specifying exactly what you need in your word. These models don't know how your formats are, what your company needs, what your report needs. So give them very detailed instructions. And then this part, which is guardrails. So if I'm talking to you or if I'm interviewing you or if I'm seeing like how good you know, if you have a good structure in your prompt, like role, instructions, guardrails, or even you can say tool sections specifically, I know that you have learned AI well. So these are the ones. So let's say that's our role, goal, and tools. So now you see did I gave company A, company B, and that automatically opened me these ports. These ports were not there earlier. And you can a- if you add, added this and say company three, you will have another one.

   6. AGAakash Gupta10:21

      Got it.

   7. MYMahesh Yadav10:21

      Now, this input I can go, get from here. I connect this, I connect this. Great. And this input, as my friend-

   8. AGAakash Gupta10:31

      So the purpose of the text fields and making it company A, company B is so that you can change out what companies are there?

   9. MYMahesh Yadav10:38

      Correct. And I will show you that you can take this from a front end as well.

   10. AGAakash Gupta10:43

       Okay. Nice.

   11. MYMahesh Yadav10:43

       Like the user can pass these, so you need not to show the user this whole backend thing or playground or this fancy thing.

   12. AGAakash Gupta10:49

       Yeah.

   13. MYMahesh Yadav10:50

       They will see a fancy website. They can put out two companies, and they can get their analysis done.

   14. AGAakash Gupta10:54

       Nice.

   15. MYMahesh Yadav10:55

       And that's why this needs to be a separate input because this you are gonna take from outside.

   16. AGAakash Gupta11:00

       Yep.

   17. MYMahesh Yadav11:00

       If you have it here, the prompt, the user is not gonna write a prompt. Maybe they just write this and press a button, and then you can do the magic behind the scenes.

   18. AGAakash Gupta11:08

       Yeah.

   19. MYMahesh Yadav11:08

       And that's why I kept them separate here. Okay. Seems like we need a tool, and as I gave you a tip earlier, we are looking for a search tool. And Tavily is an awesome tool because this is the search written for agents. What it does is it also takes out the URLs from the search, put the text out, and then give it to the agent so the agent can process lot of text or that it got from the URLs rather than just the URLs, because the URLs are not very helpful.

   20. AGAakash Gupta11:39

       Okay.

   21. MYMahesh Yadav11:39

       So Tavily is a third-party tool which need an API key. You can go to their website. It's very simple. So you can go to their website and say, obviously everybody requires you to have a login.

   22. AGAakash Gupta11:57

       [chuckles]

   23. MYMahesh Yadav11:58

       But once you log in, you can just click a button and they give you an API for free.

5. 12:03 – 15:40

   Testing Backend Agent

   1. MYMahesh Yadav12:03

      I hope this whole thing continues like this forever. So you can just take the API. Now you can see mine also, but not fully. Hopefully that w- that didn't change a lot. So now you can come hereAnd you can paste your API because your code needs that access to their API. And for this one, all you need to do is if you can just change the mode to tool mode, you need not to pass the query, and then you can connect the tool to the tool section. And now you have a fully capable agent which takes two inputs, which is competitor names. It has its instructions that goes to agent instructions. It uses this tool to do the search and gives you an awesome output. Ready to test? Maybe one last thing. It also need an OpenAI key. How you get an OpenAI key, you can go to platforms.openai. Hopefully, it doesn't land me to my prompt section and you get all to see all my prompts.

   2. AGAakash Gupta13:10

      [laughs]

   3. MYMahesh Yadav13:10

      You can go to platform.openai.

   4. AGAakash Gupta13:13

      I think I've exposed those before of my own. [laughs]

   5. MYMahesh Yadav13:16

      Oh, great. [laughs] I think then I'm not the only one.

   6. AGAakash Gupta13:20

      Uh-huh.

   7. MYMahesh Yadav13:20

      Then you go to billing, make sure you have some money left. Six dollars is enough. Five dollars is generally enough. Then go to API keys. Create a new secret key. Put your test key name, say, podcast, and select a default project, and create secret key. Once you have the key, you can copy it and come back to your flow. Replace this one with your key. And that is all. This is what it takes to create an awesome agent, which is using tool, which is using prompts, and giving you output. You want to test it?

   8. AGAakash Gupta13:54

      Can we try a different, can we try a different model though maybe instead of GPT 4.0? What if we try something else?

   9. MYMahesh Yadav14:00

      Oh, it sounds like a plan. Uh, maybe you want to change the model here.

   10. AGAakash Gupta14:05

       Yeah.

   11. MYMahesh Yadav14:05

       So there are model, and it supports all OpenAI models. And by the way, it can also take Amazon, Anthropic, and other models, but I just got the keys for OpenAI-

   12. AGAakash Gupta14:14

       Yeah

   13. MYMahesh Yadav14:14

       ... so I'm just gonna use OpenAI. But let's change this model name.

   14. AGAakash Gupta14:17

       Yeah.

   15. MYMahesh Yadav14:17

       Maybe you want to change the model name from GPT 4.0 to mini because you're doing experimentation and you don't want to be bankrupt trying a lot of things.

   16. AGAakash Gupta14:26

       Yeah.

   17. MYMahesh Yadav14:26

       So let's just, this is the smallest model with the smallest cost and the fastest results. So let's experiment with this one. If we don't get highest quality results-

   18. AGAakash Gupta14:34

       Can we use GPT-5 over there?

   19. MYMahesh Yadav14:37

       ... we can go higher. No, they don't support that yet.

   20. AGAakash Gupta14:38

       They don't support GPT-5 yet. Oh, that's sad. Okay.

   21. MYMahesh Yadav14:43

       Awesome. So now I have changed it to GPT 4.0 mini. The rest-

   22. AGAakash Gupta14:47

       And the limitation here is on Langflow side?

   23. MYMahesh Yadav14:49

       Yeah. They have to-

   24. AGAakash Gupta14:50

       Okay

   25. MYMahesh Yadav14:50

       ... expose it, right? So-

   26. AGAakash Gupta14:51

       Okay

   27. MYMahesh Yadav14:52

       ... that's the problem.

   28. AGAakash Gupta14:53

       All right.

   29. MYMahesh Yadav14:53

       That's what happens once you take abstractions, you don't... you lose something.

   30. AGAakash Gupta14:57

       Yeah. So that's maybe the one downside here.
6. 15:40 – 19:02

   Ads

   1. AGAakash Gupta15:40

      sponsor of this podcast, and also the place where Maven hosts his AI Agent Mastery course. This course has been taken by hundreds of PMs at some of the top companies in the world, like Google, Amazon, Meta, Netflix, Stripe, and many others, and it has amazing reviews. In fact, it has some of the best reviews of any Maven course I've seen, and people who, uh, in my community have taken the course have had great things to say. So if you wanna get $100 off of Mahesh's course, use code AAKASHxMAVEN. That's A-A-K-A-S-H-xMAVEN. And if you're looking for a quick place to click into that, where you don't even have to use that, go to maven.com/x/aakash. That's M-A-V-E-N.com/x/A-A-K-A-S-H. Today's episode is brought to you by Miro. Let me ask you something. How many tools are you juggling just to get a single project across the finish line? One for brainstorming, another for planning, something else for tracking tickets. That's where Miro comes in. It becomes an all-in-one collaboration workspace. Whether you're consolidating user research from several interviews, developing and synthesizing product briefs or a wireframe, or project managing development, Miro brings everyone into the same space. It's fast, intuitive, and fully loaded with features like project templates, two-way Jira sync, and integration with software like draw.io and PlantUML. Miro's AI features can be used to synthesize elements in a board to develop a ready-to-review product requirements document in seconds. If you're tired of tab overload and scattered workflows, try Miro. Head to miro.com and see why over ninety million users choose Miro to guide from idea to outcome.

   2. MYMahesh Yadav17:25

      Okay. Okay, let's go back to the model changing piece, right? So okay then, I just changed the model to GPT 4.0 mini. This seems like it's ready. Let's test it out. So to test your, uh, agent on the backend, you can just click Playground, and you just click this button called Run Flow. And you see that it's processing the input. You can see what step it is taking, how much time it is taking. By the way, when it is doing that, you can see that where it is, and you can see how much time it took across these things.

   3. AGAakash Gupta17:58

      Yep.

   4. MYMahesh Yadav17:58

      Also, you can also click here and analyze their outputs, like what was passed from here and the logging. And finally, when it is done, you can just see the results are here. I'm finished, and here is-

   5. AGAakash Gupta18:11

      Seven seconds or fourteen seconds

   6. MYMahesh Yadav18:11

      ... across these eight verticals, your results.

   7. AGAakash Gupta18:13

      Okay.

   8. MYMahesh Yadav18:14

      That's pretty cool. But maybe you want to launch this as a website in your team and don't want the world to know your secrets, that you build it in, like, five minutes and the whole world is building it.

   9. AGAakash Gupta18:24

      [laughs]

   10. MYMahesh Yadav18:25

       So maybe [chuckles] What you wanna do is you wanna create a fancy website which sends these two things, and now anybody in the world can use your tool to do this. So let me show you how you do that. So this is your backend. You can make it as complex as possible. This allows you to do RAG, multi-agents with CrewAI, or we have labs where we have gone ahead and even put agentic RAG with graph or did some fine-tuning also. For that, some other day. But today, let me just show you how can you take this and connect a lovable

7. 19:02 – 25:19

   Building Frontend with v0

   1. MYMahesh Yadav19:02

      or v0-like agent to, with Langflow so that you can do all a fancy work on the front end, but your backend is where your AI is, and that can remain separate. So for that, you need to do couple of things, so just follow along. If you click Publish, you click API Access. So they allow you to access this as a API, the whole flow. I use this curl command. So I say curl, and then I will say, okay, you see that my input-output is not here. So for that, there is this tweaks, and you can go to... Just change this to any other name, so maybe Meta and CompV to, let's say, Anthropic. Okay. Now you will see that make sure that these two inputs are there, because these are the text inputs you want your user to supply. They were not there, and you don't want this input value, so I will delete it. But let me just show you how. So you need to remove this because this is not needed, and make sure these two are here. You need security because you're calling this from outside. So what you need to do is you can just say Generate token. It gets you a bearer token, which allows you to access it from outside because some third-party app is gonna use it. So take that security, and now I'm taking a little bit in production, right? Ah, that's fun, no?

   2. AGAakash Gupta20:35

      [chuckles]

   3. MYMahesh Yadav20:36

      But, but you come here, and then this was your prompt, and maybe you copy your, uh, token, and then you copy your curl command. This command here. So I will just take this and copy it here. Awesome. I got my command. I got my curl command, and it says, "Bearer token, your token supposed to be here." So I will just replace this with my token. That's pretty much all I needed to do, I believe. I'm ready to test this from outside. And to build this end-to-end, you will need the response also. For that, you can use this tool called Postman. Anybody I know Postman is a tool which you can use to test backend APIs. If you have never used it, this is how it looks. You go to Postman. First time maybe you need to log in, then you can just say New Request. Great. Once you come here, you copy-paste the whole thing that we got from there. It automatically structures it as an API request, like what is in the header. Just make sure your key is here, and you send this request. And hopefully you should get a response. Okay, for Alpha and Brad, if you pass an input value, great. Th- this happened because I'm passing an input value and it doesn't have that. Uh, it has two same names, and I told you that you need to remove this. So now you know how to debug these also. So remove this input value-

   4. AGAakash Gupta22:14

      Mm.

   5. MYMahesh Yadav22:15

      ... that I forgot to do that.

   6. AGAakash Gupta22:17

      Mm.

   7. MYMahesh Yadav22:18

      Uh, because you have two input values, one from text input and not one for normal, and the API was only expecting one.

   8. AGAakash Gupta22:25

      Yeah.

   9. MYMahesh Yadav22:26

      So now you come here again, you paste the whole thing again, it updates it. You hit Send, and hopefully things will change for us. It takes 30 to 40 seconds to response, so hopefully we get a response.

   10. AGAakash Gupta22:39

       What is it doing in the backend here?

   11. MYMahesh Yadav22:41

       Right now, what it is doing is it's becoming like a f- application, like your web application, and it's sending this request to this guy which we built here. And you will see this request come to your playground, but you didn't type it. And something will come here and get a response and show you the response here.

   12. AGAakash Gupta23:01

       Got it.

   13. MYMahesh Yadav23:01

       So it's sending that request to Langflow and getting you that response of those two competitors which you typed, Meta and Anthropic, and trying to get you a response.

   14. AGAakash Gupta23:12

       Mm.

   15. MYMahesh Yadav23:13

       By the way, it also resolve a lot of security things on internet, so it will pass the bearer token, it will make a handshake, and here it is. And it says, "Here is a comparison of Me- Meta, Anthropic-based this," and it gives you a bunch of values because that's how behind the scenes computer talk to computer, which is lot of JSON, lot of data. Maybe you wanna do fancy things with it. So what you can-- what you d- wanna do next is just copy this and paste it into what we were using. So just use this. So this was your command, and below this, just paste the response you got. So far with me?

   16. AGAakash Gupta23:54

       Yeah.

   17. MYMahesh Yadav23:54

       So I got this command I created. I got the response. That is all you need to do. So once you know what command to send and what the response is, you can go to a tool like v0. And then I, uh, I'm gonna give you a prompt, which is here. You write, take this prompt and say, "You know what? I am cool." It takes me to be cool. [chuckles] Here is a, here is your prompt, which is-So let me show you the prompt. So this is a front-end building tool, and here you are saying, "Hey, build me a modern responsive landing page for a competitor comparison tool using provided APIs." It takes two inputs, company A and B, and compare button. It calls APIs, resolves some issues like CORS, which is cross-resource resolution. If one party calls the other party, there is a normal error you get. I put it as a error handling. I also tell it to do error handling, and that's pretty much it. And here, you need to response your com-command that you send. Okay?

   18. AGAakash Gupta25:07

       Yeah.

   19. MYMahesh Yadav25:07

       So you remove this, you take it from here, the command that you just sent. So,

8. 25:19 – 30:17

   AI Agent vs Regular AI Product

   1. MYMahesh Yadav25:20

      so you tell it right, "Hey, this is how we're gonna call the API, and here is the response format." What is, what is this allowing v0 to do is it now can automatically figure out how to send this request, and g- once it gets the response, it will show you in a very nice format, but it need to know how the server is gonna respond. In past, a developer need to read maybe 20 API documents to get here, but now all you are doing is copy-pasting the response, and that is it. You can send this request right now, and hopefully in two minutes your front end will be ready, and you will be ready to publish your app. So that's pretty much all it takes these days to build end-to-end apps with a solid back end with the AI and a front end, which is a website which I will publish on my site, say maheshyadav/competitor-analysis. And then as you guys respond, you say, "I want this step. I want that step," I can keep iterating and keep building and eventually build the best competitive analysis tool the world has ever seen.

   2. AGAakash Gupta26:32

      [chuckles] Okay. So right now v0 is working. It's just getting its design expira- inspiration.

   3. MYMahesh Yadav26:37

      Correct.

   4. AGAakash Gupta26:37

      It's exploring its code base structure. It's-- You can even click into the code base, which you're doing here. So you can see it always uses, uh, pretty much always uses Tailwind CSS as its default. So it went ahead and did that. It's developing its color palette, just a bunch of basic CSS stuff right now. Then it's doing its layout. So in this case, like up until this point, we have armed v0 with a really good prompt, and so we'll include that prompt for folks in the newsletter article so that you can just copy-paste exactly what Mahesh used. But I think that what's probably most important there is it wasn't, um, just a simple one-line prompt, right? You had a lot of specific points you included. Um, while we're waiting for this to outline, can we just cover what were those like four or five key points in that prompt again?

   5. MYMahesh Yadav27:25

      Sure. So that prompt, by the way, I went ahead and did something more for Aakash-

   6. AGAakash Gupta27:31

      Uh-huh

   7. MYMahesh Yadav27:32

      ... and his listeners. So what I did is, because it could be audio, I have created this whole lab with screenshots and everything.

   8. AGAakash Gupta27:39

      Nice.

   9. MYMahesh Yadav27:39

      And we can all-- We are open sourcing this, so you will get this link. This allows you to step by step follow what I am doing. And here you can see how I created my agent, how I created my back end, then how I am going and creating this Postman thing, getting a response. And then I show you this prompt on v0. We also introduce you to v0. But that's the prompt we were using, and the prompt had this like what is the task that needs to happen?

   10. AGAakash Gupta28:07

       Task.

   11. MYMahesh Yadav28:07

       What are the requirements?

   12. AGAakash Gupta28:08

       Requirements.

   13. MYMahesh Yadav28:08

       And how to get the resources to es-establish this.

   14. AGAakash Gupta28:12

       Resources. So it's like a three-step prompt. Okay, perfect.

   15. MYMahesh Yadav28:15

       Correct.

   16. AGAakash Gupta28:16

       All right.

   17. MYMahesh Yadav28:16

       And I can-- I put like lot of guardrails here. If you look inside, I said that, "Hey, make sure you're doing error handling."

   18. AGAakash Gupta28:23

       Right.

   19. MYMahesh Yadav28:23

       Because automatically it doesn't do error handling. I said that I need smooth animation and transitions, and I said that any error that you might have got before, you can just keep adding to your prompts-

   20. AGAakash Gupta28:34

       Yeah

   21. MYMahesh Yadav28:34

       ... so it automatically handles them.

   22. AGAakash Gupta28:37

       Yep. That's the art of prompt engineering is to go ahead and do that error analysis and then update your prompt.

   23. MYMahesh Yadav28:44

       Finally, I think it's getting there. It's generating its final TypeScripts. So hopefully we will have it. It got its cards, set company A, B. It's fun to see all these live.

   24. AGAakash Gupta28:58

       It's so much faster than you would've done it as a human.

   25. MYMahesh Yadav29:00

       Oh, man. [chuckles] I would have not done it. I'm very bad in bad, uh, front-end code. Uh-

   26. AGAakash Gupta29:05

       Oh.

   27. MYMahesh Yadav29:05

       These are like came as a rescue.

   28. AGAakash Gupta29:06

       I did that like back in 2008. It hasn't actually changed very much. [chuckles]

   29. MYMahesh Yadav29:09

       Oh, really?

   30. AGAakash Gupta29:10

       Yeah.
9. 30:17 – 31:48

   Vibe Coding Interviews at FAANG

   1. MYMahesh Yadav30:17

      and let's say Microsoft.

   2. AGAakash Gupta30:21

      Or what if we give it a harder one? Let's do Mistral. M-I-S-T-R-A-L.

   3. MYMahesh Yadav30:25

      Sure.I wish. [laughs] Uh, like this is like if you would have written the code or if you are supposed to write the code, you worry about these things a lot, but now who cares? All we have to do is just modify the prompt, so try anything. Now you have full control, right? And at this point, it's sending that request out, and it might take couple of iterat-iterations for this to get it working. It might fail, and I want to show, like, what you can do when it fails. But at the same time, you s- when you send this request, what happens here is it's sending this request here, and you will see that it will come here, it will try to get... You see that?

   4. AGAakash Gupta31:04

      Oh, okay. We get to see it.

   5. MYMahesh Yadav31:04

      So this was the last one from Meta and Anthropic, but it will, in a second you will see. So this is the one we sent from our Langflow.

   6. AGAakash Gupta31:12

      Yeah.

   7. MYMahesh Yadav31:13

      I changed it to Meta to Anthropic.

   8. AGAakash Gupta31:15

      Yeah.

   9. MYMahesh Yadav31:15

      And maybe in few seconds you will see that coming too.

   10. AGAakash Gupta31:19

       Mistral, yeah. We're looking for that new one.

   11. MYMahesh Yadav31:21

       Yeah. So that's coming, and hopefully if it is able to send the request, generally it fails in the first-

   12. AGAakash Gupta31:29

       If there's no errors. [laughs]

   13. MYMahesh Yadav31:30

       Yeah.

   14. AGAakash Gupta31:32

       It, like-

   15. MYMahesh Yadav31:32

       Okay

   16. AGAakash Gupta31:32

       ... stopped showing the loading.

   17. MYMahesh Yadav31:34

       No, no, it came, like-

   18. AGAakash Gupta31:35

       Oh, it came. It just didn't-

   19. MYMahesh Yadav31:35

       ... Google versus Mistral. Google, it started in 2023. Arthur Menache is the emerging player in AI, venture-backed, privately held. Make sense?

   20. AGAakash Gupta31:45

       Today's episode is brought to you by the experimentation

10. 31:48 – 43:45

    Ads

    1. AGAakash Gupta31:48

       platform Kameleoon. Nine out of 10 companies that see themselves as industry leaders and expect to grow this year say experimentation is critical to their business, but most companies still fail at it. Why? Because most experiments require too much developer involvement. Kameleoon handles experimentation differently. It enables product and growth teams to create and test prototypes in minutes with prompt-based experimentation. You describe what you want, Kameleoon builds a variation of your webpage, lets you target a cohort of users, choose KPIs, and runs the experiment for you. Prompt-based experimentation makes what used to take days of developer time turn into minutes. Try prompt-based experimentation on your own web apps. Visit kameleoon.com/prompt to join the waitlist. That's K-A-M-E-L-E-O-O-N.com/prompt. AI evals are one of the most important skills for PMs, and I know you know they matter. The question is, are you doing them right? Most teams are winging it with basic metrics and hoping for the best. Meanwhile, the teams that actually ship reliable AI, they've cracked the code on systematic evaluation. Today's episode is brought to you by the AI Evals for Engineers and PMs course by Hamel Hussein and Shreya Shankar. This live Maven course will teach you the battle-tested frameworks from Hamel and Shreya, who are the engineers behind GitHub Copilot's evaluation system and 25-plus production AI implementations. Four weeks, live instruction, next cohort starts July 21st. Start shipping AI that actually works. Enroll at maven.com with my code AG-PRODUCT-GROWTH for over $800 off. That's A-G-P-R-O-D-U-C-T-G-R-O-W-T-H. Today's episode is brought to you by Amplitude. Replays of mobile user engagement are critical to building better products and experiences, but many session replay tools don't capture the full picture. Some tools take screenshots every second, leading to choppy replays and high storage costs from enormous capture sizes. Others use wireframes, but key moments go missing, creating gaps in your understanding. Neither approach gives you a truly mobile experience. Amplitude does things differently. Their mobile replays capture the full experience, every tap, every scroll, and every gesture, with no lag and no performance hit. It's the most accurate way to understand mobile behavior. See the full story with Amplitude. [lip smack] So now we've got it. So I guess let's publish this, right? How do we go to the final step?

    2. MYMahesh Yadav34:22

       Yeah. So you can go and say Publish to production. And it's deploying it. Here you can also provide when it does deploy. You can see, so Vercel is the backer. Vercel was the de facto standard before this to publish your prototypes anyways. So it's putting those. It will take a second, but now I will give you a link. On Vercel it will be right there, but then you can go to Settings, and you can also provide your own domain name, like maheshcompetitoranalysis.com. But that's, you need to go and buy that domain first. But in a second, you will be able to get a URL here that you can use or send it to your friends.

    3. AGAakash Gupta35:00

       Okay. So it'll host it for you.

    4. MYMahesh Yadav35:05

       Yeah. It hosts it. It will scale it also if you do, but it will have a Vercel in the end because it's free, so they at least want to make sure that they are getting marketing money.

    5. AGAakash Gupta35:16

       Yeah.

    6. MYMahesh Yadav35:16

       So, [chuckles] uh, so this is your app, and you can open on your screen. So I can go to any right now, and I can just say, "Dude, here it is, and this app is live."

    7. AGAakash Gupta35:27

       Nice.

    8. MYMahesh Yadav35:27

       And you can try anything else. Maybe now you want to say DeepSeek.

    9. AGAakash Gupta35:31

       Yeah, DeepSeek, and maybe like Amanus, A-M-A-N-U-S.

    10. MYMahesh Yadav35:36

        These are not companies, so hopefully we get more interesting things. [laughs]

    11. AGAakash Gupta35:41

        Yeah, I think these are products. [laughs]

    12. MYMahesh Yadav35:43

        Yeah. Well, let's see. It's always fun. Anyways, you're doing market research, right? And hopefully some of this will be weird because maybe it will figure out the parent companies and compare them. So that's the idea.

    13. AGAakash Gupta35:56

        Yeah, if it does that'd be-

    14. MYMahesh Yadav35:56

        So if you go to Data Stacks here, if you just refresh it, you will find, like, in couple of seconds you can get a new information here too. Sometimes it just takes a while for it to just trigger it, especially if you're triggering it from outside this playground, doesn't stay in sync. But I've always seen it eventually comes here.

    15. AGAakash Gupta36:15

        There it is. Yep, here's Mistral. So Mistral finally came through.

    16. MYMahesh Yadav36:18

        Yeah.

    17. AGAakash Gupta36:19

        Okay.

    18. MYMahesh Yadav36:19

        So now next time you will come and see, you will have it. But I don't like that star star, so let's change that-

    19. AGAakash Gupta36:25

        Yeah

    20. MYMahesh Yadav36:26

        ... and let's change the little formatting of our web app also. I will show you a couple of things on v0, which I always do. So-Another thing that you can do is, okay, you have published it, it looks good, you can try it. Okay, there is an HTTP error, and you can fix it. So sometimes what happens is there is a delay, and you get this five hundred error because that endpoint is not responding. In those cases, just try again, and it will work. Obviously, when you're gonna put it to scale, you want that code to be hosted only on your server and s- make sure your response times are awesome. But for free, I think we are good to go. So if you go back here, what you can do is... This is not looking good, so I'm like, "Hey, what happened here? Fix this." So you can just take a screenshot of anything, and you can just say, "Hey, fix this part and make it bold and blue really."

    21. AGAakash Gupta37:29

        And remove the stars-

    22. MYMahesh Yadav37:29

        And just paste it

    23. AGAakash Gupta37:30

        ... right?

    24. MYMahesh Yadav37:31

        Yeah. And I think it will figure out what is the fix here, right? So I will say, or you want to say explicit. Let's try without saying explicit, and then you will see.

    25. AGAakash Gupta37:41

        Okay.

    26. MYMahesh Yadav37:41

        So I think I just-- If these are intelligent agent, they should be intelligent. [chuckles]

    27. AGAakash Gupta37:46

        This is 4o-

    28. MYMahesh Yadav37:47

        So-

    29. AGAakash Gupta37:47

        No, this isn't 4o mini. This is whatever v0 has going on in the background-

    30. MYMahesh Yadav37:50

        Correct, correct
11. 43:45 – 50:51

    History of AI Agents

    1. AGAakash Gupta43:45

       this place? What's the architecture of an AI agent?

    2. MYMahesh Yadav43:48

       Actually, I have a slide for that. So [chuckles] let me just walk you through that. So a short history of agentic AI. Let's talk about that because I was lucky enough to very, to start or been working in agentic AI for a while, so I created this one for all the audience that are looking at it. So on x-axis, I have time. On y-axis, I have value. And this is the first time you see that everybody can remember that moment when you first time heard about ChatGPT. It was such a compelling proposition that you took your data to the product. That has never happened in AI before. By the way, I was trying to sell AI-based products seven or six years before ChatGPT, and we have to sign hundreds of NDAs, goes through corporate loops to get people data. But this was one product where people actually took their data and copy-pasted in ChatGPT and said, "Analyze it for me." So that was the moment where we started seeing success in chatbots. This is where we solved the Q&A problem. And in 2023, what happened is we said, "Okay, you need now to take your data to ChatGPT." The big giants woke up, like Amazon, Google or Microsoft, and they launched some version of Copilot in their products. So now you can have the product, you can right swipe and ask questions or do things which were not possible before. And then 2024 we made, reintroduced agents. I was lucky enough to work on the first framework to build agents at AWS, and there we said, "Okay, Q&A is cool, but what about I let you connect tools or make changes in the world, or change the state of the world by calling APIs, doing search, writing code?" All these tools, if you connect to these Q&A machines, can they start doing things? And that's agents. And then in 2025, we are making them multi-agent and multimodal. What does that mean? That it only not take text, but can take images, which I was showing you on v0, that how can I just paste a picture and say, "Hey, make it look like Gartner's website." Or I can just send my same through audio also. So we are making it multimodal. Also, we are saying not a single agent is enough, maybe connect many agents to do complex things. One good example of multi-agent is the AI coding agents, which we were showing again in v0, that this is not a single agent. One agent is thinking, the other agent is writing the code, other agent is testing, and all that combined are able to do complex things. That's what is so exciting. So we have gone from simple chatbots, which were amazing, can summarize, can generate forms and all, to fully blown multi-agents which can write code like all of us do, create web apps or do PhD level research for us. So that's our journey so far.

    3. AGAakash Gupta46:55

       So that covers the history of agentic AI. Now let's talk about these vibe coding interviews, these AI coding interviews. We just demoed an end-to-end workflow for people. A lot of people are gonna need to execute these in interviews because places like Google are bringing on these interviews. You worked at Google, you helped them write some of their first AI guidelines. What do FAANG companies like Google look for in these interviews?

    4. MYMahesh Yadav47:17

       That's a great question. And AI is new, right? So what we are looking for people is that have actually built in AI, because what we want to avoid is people who are setting the expectation too high or setting them too low. And we don't really care. We understand that you have not work in production, but have you actually built a end-to-end workflow and maybe ten people tried it? How many iterations of that? So we are looking for builders and not people who know just frameworks anymore. So that's the most exciting part for people like me. [chuckles] Second thing, we are looking for people who have done interactions in AI specific world. Let me explain. What that world looks for is, do you have sense of how to interact or get your data or handle data responsibly? Can you understand how these models are designed and which models are good at what? Can you evaluate an agent or AI, or can you work in iterations? And third, I, uh, earlier talked about, which is how much things have you achieved at scale? Not in AI, maybe with some other example. You can bring your cloud scale or if you pivoted in mobile. But we are looking for people who have seen one technology transiti- transaction going from one change, which is big enough, and how they handled it. And if you have two of these, I think you have a really good chance to get into any of the main companies because they are desperate to, to be honest.

    5. AGAakash Gupta48:44

       So this PM interview type that has spotted up, this vibe coding interview, they often ask you to vibe code live in the interview. What's the right way to handle this interview type? What's the right framework and structure to ace it?

    6. MYMahesh Yadav48:56

       Yeah, I think that's a very good question, and I thinkIt's been thought perceived differently, so let me take you- tell you my take on this. So I think when we are asking you to vibe code or when anybody asks you to do vibe coding, the requirements has not changed. We are not looking for your technical skills of how good you write code or do you understand what is written. What we are looking for is what are you adding in your... It's just a easy way for us to create something together and help me judge how you think as a product manager. So when you're vibe coding, do not forget your product principle. That's my lesson number one. Second thing, please play with these tools. They're not very hard. So what I'm looking for there is like the prompt I showed. Do you structure your prompt? And that tells me that, yes, you have done some in- interactions with AI, and second, you understand the limitations of these models that they can't take mumbo jumbo instructions correctly. So this is the second thing I think people are looking for vibe coding. And third is, more than you just show me how cool this thing is, you show me these iterations. That tells me what is your taste. So I'm looking for two things, to be honest. I'm looking for taste and can you improve with iterations. Can you evaluate and then improve it? So show me maybe three things. One, show me how, how you think and show me in your prompt how you think. Second, once you build it, show me what kind of user insights you are bringing in your second iteration. And hopefully third iteration, show me some feedback loops that I will make for this, which shows me that you have a sense of data and AI in you. And I think that could be a very good interview or three-step framework that people are looking for when we ask you to vibe code. So just take it like a canvas, I would say, to show your PM skills.

    7. AGAakash Gupta50:48

       One of the things that we talked about in the pre-show

12. 50:51 – 53:58

    Cart Before Horse Development

    1. AGAakash Gupta50:51

       conversations we were having for folks was this idea of cart before the horse development. I loved your take on this. What is cart before the horse development?

    2. MYMahesh Yadav50:58

       I think now all of, all of the PMs can hate me after this. But [chuckles] but the idea is this, right? The idea is that PMs just every room they go to, they stop engineers saying that, "Hey, this is us, and we will tell you, uh, don't put cart before the horse." But in AI, that's changing because that's the whole idea of having a PM because they decide like where the horse goes and then build the cart. Uh, but to be honest, I think this is the first time we are actually in a position to change that idea a little bit. Because what's happening is three things. One is the cost of prototyping has gone 100X lowers- lower in last two years itself. So building something, as you saw, is very cheap and anybody can do that. That's number one. Second, the customers really don't know what to expect out of AI. And third is there's this idea that so many possibilities, so many ideas, so many things are possible, and if you're gonna do s- six months of research to figure out which problem to solve, maybe somebody who did something else will solve it before that. So this FOMO factor. So these are the three changes, and what it is doing to our industry or what is a... It's changing the expectation from PM from saying, "Hey, I will take this..." So old method is I take this problem, I do a research three months, I write this awesome piece called PRD. In another two months, I get these approvals in two months, and then I hand it over to engineering, and hopefully I will show up when business needs me or when it is done in another six months and we launch one product or one feature every year and I get promoted. That's the old world. The new world is the PM actually talks to customer two weeks, three weeks, creates a prototype, show them what's possible, and then iterate on the prototype maybe for two or three weeks. Then write a very small PRD with detailed user experience that this is what happens on this click, this is what we should show, this is what we should show. A clear user experience with ideally some prompts and then evaluations, which is how we're gonna evaluate what is the quality expectation. And with these three things, they can hand it over to engineering. Engineer makes it possible and make it production ready. And then you iterate every three months, and in a year you have built a product and actually made it make money and as, and made it what exactly you could never think or nobody will ever tell if you have not done that. And that's the new way of doing AI PMing or PMing. And in this case, you're building the cart first and then working with the customers to figure out where the horses will go. And that's the idea of cart first AI product development.

    3. AGAakash Gupta53:54

       Love that. So I wanna shift a little bit into how you

13. 53:58 – 1:05:10

    Cracking FAANG Interviews

    1. AGAakash Gupta53:58

       broke into FAANG, because a lot of people, they wanna break into these companies. They want to get this 300,000-plus PM AI Agents job. So let's start with your own story. How did you break into AI at Microsoft? I believe this was back in 2016.

    2. MYMahesh Yadav54:13

       Yeah. No, I think very good question, right? So I think first is having an intuition for a new technology. I had that. I was working in Azure IoT, and I thought maybe this idea that you can, uh, instead of hard coding or doing logic-based programming, this whole ML thing will be a cool thing. But as you may all be feeling, that it's not easy to enter. There are barriers of terms like, you know, gradient descent, local minima, math. You need to be this research-y scientist. It was even worse in 2016. And for me, I just wanted it so desperately that I started hanging out around Microsoft Research campus. I offered my services on Friday, Saturday for free. And, uh, then I came up with one idea of building a Vision AI dev kit, which was a camera that can inference on-Edge but train in cloud. And I offered that to our Azure machine learning team, and I was a dev then, and they said, "You need to test." And in Microsoft, if they tell you to test, it's not a good deal.

    3. AGAakash Gupta55:20

       [chuckles]

    4. MYMahesh Yadav55:21

       But I took anything. I said, "Yeah, I will test it." And in three months, some people dropped something. I took on that code, then code, and eventually I became the TL for that. I shipped that product end to end, and that was my story. I was the guy who actually built Vision AI Dev Kit and shipped it end to end to production. Satya took it in our 2017 or 2018 Build on stage. And after that, it just a story from there. I knew a lot of things, like I knew how chips acceleration work. I knew different architectures. I knew how to train them in cloud. I knew how to optimize them to run on small hardwares. And after that it just iterations and I think after that, it just becomes like once... It's like a marathon where a lot of people run with you, and if you get to do the first good project, you get a good lead, and from there you can run the marathon correctly or without, like, running into people or getting stopped every five minutes because nobody's moving in front of you. So my two tips out of my own experience is be ruthless and passionate. Take anything that's coming in ML and AI your way, or create your own paths. And once you have them, don't take them for granted. Work twice or three times harder once you have those opportunities, and I, I'm pretty sure it will change your life. It has changed mine, since I was able to double my salary every other year after that.

    5. AGAakash Gupta56:51

       Wow, that's insane. Talk us through how you made it from dev to PM.

    6. MYMahesh Yadav56:56

       You have really good questions for me. [chuckles] So, so I was, uh, when I was doing this, uh, Vision AI kit, I also realized this another insight, which was in AI, I believe that building things will get easier and easier. But what to build and how to work closely with customers to build the right thing would be an skill or would be a s- thing that will save a lot of time for organizations or developer time. And to be honest, I hated, like, what PMs did because they could just tell me, and then I have to spend six months, and I couldn't, like, question a lot. So but I found them, like, little vague.

    7. AGAakash Gupta57:35

       [chuckles]

    8. MYMahesh Yadav57:35

       So I don't never wanted to get into PMs, but it, the stakes were so high for me that I did like two iterations and things didn't work out, and I don't want to blame anyone. So I said, "Okay, let me figure out." And my God, it's so hard to figure out what's the right thing to build. [chuckles]

    9. AGAakash Gupta57:54

       [chuckles]

    10. MYMahesh Yadav57:54

        It's super easy when somebody tells you what to build to build it. But figuring out what is the right thing to build, especially for a large company, when the stakes are higher and your role is little, like you get into L7 levels, uh, you're playing with at least 300 to 400 people's life, and you're almost playing with $50 to $60 million worth of decisions you are making. So why I started? I started out because I wanted to make sure that we are building the right things, and I think what I loved about it is learning how to build the right things, which I'm still trying to figure out.

    11. AGAakash Gupta58:30

        How do you architect that internal transfer, though? Some people, they just get stuck in the engineering ladder. They can't make it into the product management ladder.

    12. MYMahesh Yadav58:37

        Oh, yeah. I think that's a, that's a very good question. Uh, and I seen that. I seen a lot of my friends who actually wanted to become PMs, but nobody, like, made it. Uh, and I was trying to stay back, and the PM manager actually said, "Mahesh, I want you in my team. Can you join my team?" And I said, "No, I can't join your team because my visa has this, you know, problem," all, all the thing that happened in US. Uh, but he kept pushing, insisting on me. And, uh, why? To answer your question, because I was super passionate about customers. I was working backward from there. I was there to solve the problems. I could prototype real quickly, and I cared about business. I cared about making money for the company. I cared about what ma- will matter when all AI dust will settle, and that was 2018. This is not like now it's just a storm, not dust. [chuckles] But, uh, I cared about these three things, which is, hey, are we building the right thing for our customers? Is it gonna make money for us? And who-- how can we bring people together to build it? And I think these three skills were essential for being a successful PM. And if you show that to people, if you do that without given that, it does require time, effort, and energy, to be honest, because I need to still code. But I went ahead and did that anyways because I wanted to code what will matter, not code what will things won't matter. So I think those three things, and if you are struggling, start doing these three things, uh, which means writing a lot about what is the problem, how many... Just don't write PRDs because your PM friends will get offended.

    13. AGAakash Gupta1:00:20

        [laughs]

    14. MYMahesh Yadav1:00:21

        But write your customer insight. So I had this customer insight newsletter I started.

    15. AGAakash Gupta1:00:25

        Oh, wow.

    16. MYMahesh Yadav1:00:26

        And because I was debugging each customer problem with Vision AI Dev Kit, I will write, uh, as if like I talked to them, this is their problem, this is the model they are using, this is why this fails, this is what our system have gaps, this is what AI as a whole has gaps, and that's why I just patched it right now, but this could be at the next thing for us. One example of that is ONNX. Like we build it, uh, means, uh, you know, you don't know what to build because you're cool, you're a new PM. So we said, "You know what? We're gonna kill NVIDIA, and we're gonna build it on Intel chips." So when I went to Azure Stack, uh, which is this idea that you can run AI on premise-Rather than sending everything to cloud. That time we were dealing with government contracts, and we took Intel FPGA chips for that design, not NVIDIA. And then I learned quickly that these architectures like, uh, that time we were doing vision models, so MobileNet, YOLO, and each time there was a new architecture comes, Intel can't catch up, and it works out of the box in NVIDIA. So I wrote these, and then we came up with this idea of ONNX, which is Open Neural Netw- Network Exchange. And the idea is that Microsoft build a layer which allowed any architecture to automatically done accelerated computing across any chips. So you can train a model in TensorFlow or PyTorch. You can convert it into ONNX that Microsoft provides you with o- open source operating sy- open source framework, and then this model can run on any hardware accelerated, and you or the cloud provider need not to do any specific thing. So once you start making these kind of contributions, you can actually... Nobody will, like, stop you. Either you're gonna become a PM, or you're gonna become a GM through engineering ladder because people want people who are taking broader impact and broader challenges than what is given to them.

    17. AGAakash Gupta1:02:42

        Love it.

    18. MYMahesh Yadav1:02:42

        Small prompt, long response.

    19. AGAakash Gupta1:02:44

        [chuckles] And then after that, you worked at so many of the other FAANG response, FAANG places, Meta, Amazon, Google. How are you cracking all of these FAANG interviews in PM?

    20. MYMahesh Yadav1:02:55

        Oh, I think once it was very hard, to be honest, right? I, I was born in Microsoft. I was supposed to die in Microsoft. It's such a... Microsoft is such a family, right? Because it's especially on this side of the, uh, Lake Washington. It's on this side, Seattle on the other side, in Seattle. So in Redmond, if you live, you live in Microsoft bubble, which is you go to restaurants where all Microsoft people, you go to the gym with Microsoft people, you all your friends and all your kids' friends are Microsoft, and it's just hard to live that, uh, leave that place. Uh, so first time when Facebook reached, uh, reached out for, for me, I thought at that time I was working on inference, which is like, how can you run these models? But I wanted to go into training, and they had an awesome opportunity for me. So that time it took a lot of time, but after that, I think, to be honest, AI has done a great deal for me. Uh, I had this except, uh, a different skill set, which was in demand across these like four or five years where I've done these transitions. So I would say that played one role. Uh, second role is that I was able to just articulate well my stories, show scale, show how I handle ambiguity, and third, this never giving up attitude, which I think these companies are hungry for. Because once you enter these companies, the people there are just too hard to work with, and things don't move fast. And if you can say that I'm gonna come and make things go fast, these PM managers are hungry for you, especially in AI, because the everybody had that FOMO pressure, and this guy knows what to be built, what, what we can build, and he has done an iteration outside, and he's hungry. He is impatient. These three, I think, just made it so easy for me.

    21. AGAakash Gupta1:04:55

        Okay. And how did you get so good at the PM interview, though? Like, the PM interview, case interviews, what was your approach to that?

    22. MYMahesh Yadav1:05:01

        I think I, I watched all your series, all your, uh, read all your blogs. [laughs]

    23. AGAakash Gupta1:05:07

        Awesome. That's amazing.

    24. MYMahesh Yadav1:05:09

        Uh, yeah, I think, uh,

14. 1:05:10 – 1:11:12

    Essential AI Tools for PMs

    1. MYMahesh Yadav1:05:10

       I think just, uh, I... No, I truly, I followed you and others, and I was able to, like, at least keep, like, what's out there. There is a lot of good content or summary, but I did-- I generally, I don't stop in LinkedIn. I take what you guys fo- post, and then I look at the links, and I read them. Like, I was reading, uh, there's another guy named Lenny, I think, uh, and he posted this AI jobs, and then I was trying to figure out, okay, what's the trends look like? Is this gonna be the right thing for us, for the economy? And then I read more, like, five reports. So I think staying on top of this, that's one thing I would suggest, and don't do your own research. Let people like Akash, Lenny work for you, uh, so that you can get from an abstract form and then scale it to your own world. So that's one thing I would suggest. Second, I think just, uh, having stories where you show these three things, which is scale, ambiguity, and what is your impact on business, and show it, say it succinctly, and let the interviewer run the, run your, uh, interview, but land these ideas. And third is just having a opinionated view on things, I think was helpful for me. Like, don't be a pushover. If they push you, just, just, just say... It means I'm very respectful, and I'm very, like, I listen to them, but I try to say why I thought. And you can go with me, like, three levels down, and I will be able to tell you that my fundamentals, my, my model of the world is not very bad.

    2. AGAakash Gupta1:06:49

       Mm-hmm.

    3. MYMahesh Yadav1:06:50

       And then you can say, "Okay, maybe in this case, he might be right."

    4. AGAakash Gupta1:06:54

       And going back to your very first, at the very beginning of the conversation, you've developed, "I can build AI. I can build it at scale." And so that's, those foundational experiences allow you, when they're pushing you in the interviews, to just go-

    5. MYMahesh Yadav1:07:08

       Yeah

    6. AGAakash Gupta1:07:08

       ... three levels deep, where other people might not be able to. So you worked at all these companies. How would you compare them? How would you compare Microsoft, Amazon, Meta, and Google? Which one is best for which people?

    7. MYMahesh Yadav1:07:19

       I think that's a great question. Uh, it's hard, but, uh, let me give it a shot. So what happens is, yeah, I will tell you what, like, I loved about each company, okay? So let's start with Microsoft.So Microsoft is a dreamland for people who want to go build without caring how it will make business. [laughs] I don't know, it just ingrained in Microsoft to just innovate without figuring out how will we make money. Uh, that's... The opposite of that is Amazon. If you are working in AWS specifically, you have to be very clear how you're gonna make money because if you want to be a PM, if you want to build a lot and try taking a lot of shots, then go to Microsoft. If you want to be a PM who owns profit and loss statements, who actually cares a lot about business and ready to die for that $1 thing and execute at speed of light, go to Amazon or AWS specifically. Specifically AWS AI teams, uh, so much fun there for that one aspect, but they won't let you innovate a lot. That's the downside. That brings you to Meta. [laughs] Meta is a company where you will go if you want to innovate a lot with the right amount of people. Uh, here, one thing that I still remember is the PLs there are world-class.

    8. AGAakash Gupta1:08:45

       Oh.

    9. MYMahesh Yadav1:08:45

       And they will come to you and say, "Tell me what you want me to build." And you say, "Dude, I'm not sure. I have these, like three ideas, and all are like 20%. I'm sure that they are successful. They will be successful based on interactions I had." He said, "Can you give me one? And I will build it, and then you can launch it. And then if it doesn't succeed, I won't even tell my manager." And they will run these expe- uh, they will take on these experiments. They will build it over the weekend. I will, uh, be able to launch it on some sample set, and if it works, all they want you to f- from you is create a post and publish it, and then they will keep coming to you. But if you can't give them good ideas, if you fail three times back-to-back, they're going to the other PM, and they never come back to you, and that's your death in AI world if you're working for Meta. So Meta is a place where you want to have good, uh, iterations and working with great people to do the next level thing, um, and also never worry about money inside Meta. They have a lot of like... I felt, at least I was there before, like this whole, you know, squeeze and everything. So it's a very rich company, and it's an awesome company inside. Outside, people think badly about, you know, whatever they do. But inside, it's a very homely and awesome company, and you go there to build latest technology, whatever it could be. If Metaverse, if tomorrow quantum becomes a thing, they will be able to catch up. So if you want to stay on the edge of things, go to Meta. And last one, I will give you Google. Google is an awesome company for user experience. Uh, they just ob- are obsessed on user experience. If you want to learn how you serve anything on user experience, there's so many experts there, like what a copy should look like, where a button should go, what happens from this screens to this screen, and how can we create magic between this experience even if there is a five-second delay. They are obsessed on that. And second is they give you a lot of time. So it's like Microsoft, but they care about money and user experience. So if you add little bit of money, less money, not like crazy like Amazon.

    10. AGAakash Gupta1:10:59

        Amazon. [laughs]

    11. MYMahesh Yadav1:10:59

        So little bit of, little bit money, but a lot about user experience. So I think you wanna go work for Google.

    12. AGAakash Gupta1:11:07

        Okay.

    13. MYMahesh Yadav1:11:07

        I hope that helps.

    14. AGAakash Gupta1:11:08

        So we have been talking so much about how to

15. 1:11:12 – 1:15:56

    AI Agents Jobs

    1. AGAakash Gupta1:11:12

       break into these FAANG companies, and we started with these agent PM jobs. How many AI Agent PM jobs are there really out there?

    2. MYMahesh Yadav1:11:21

       That's a great question. Let me just show you. I got this data from this c- uh, site called TrueUp. They do, did this analysis, and they show you all the data that, that, hey, how these jobs are. By the way, they were also cool to show you who's getting paid the most. So I felt a little bit FOMO that why I'm not there yet.

    3. AGAakash Gupta1:11:42

       [laughs]

    4. MYMahesh Yadav1:11:42

       But, uh, well, let me quickly show you what this data looks like. So this is the data on jobs in agentic AI or AI specifically. So these are all the jobs postings that are open. And don't look at the numbers because, you know, all these people who scrape these jobs have some problems. But look at the trend, right? You're looking at a lot of jobs in PM, then dropped, and these are all the PM jobs. If you go to the next slides, you're looking at all the AI jobs, and now you see like where the growth is going or coming from. So this is TrueUp, and you're looking at 25K AI jobs. And then this is the last one where Lenny did, did this analysis working with, again, TrueUp. And here it shows that if there are new jobs, the likelihood of that job in AI is double than the likelihood of that jo- that job in PM space.

    5. AGAakash Gupta1:12:37

       Oh.

    6. MYMahesh Yadav1:12:37

       So this is like all PM is down here, and I'm just looking at delta because these numbers are not, like proportional. But if you look at this, that, just the comparison on that scale, this is growing two to three X more than what your traditional PM jobs are growing.

    7. AGAakash Gupta1:12:55

       What is the salary and compensation for these AI Agent PM roles?

    8. MYMahesh Yadav1:12:59

       That's, [laughs] that's a really good question. I think, uh, easily like you're looking at level six or level seven roles, uh, at MAANG companies. You can expect anything from seven fifty and above. I think that's a normal base these days. And if you are looking for level eight and above, you're looking at one point two to one point five million. Uh, that could be a easy, uh, and I'm saying total comp, like your joining bonus plus your stock, and even if the stock remains same, that's the comp these days across these companies. And-Most of the startups are trying to give even better. So if you're going to OpenAI, you can just at level six, level seven, you can even expect 900K. Netflix recently posted a job, I think they posted some 790 or 950 across these jobs, and that's normal for top 10 AI companies. And I think that's why people are excited about or people spend so much time and years preparing for that. Uh, so I think that's a good number.

    9. AGAakash Gupta1:14:00

       Insane numbers. Okay. And can anyone become an AI PM?

    10. MYMahesh Yadav1:14:05

        I think so. Uh, I think don't expect to become an AI PM if you have never worked in AI before in six months or six weeks, to be honest. It might take you a year to get to this level where you can start interviewing at, like, top companies and get to 700 plus salary. Uh, but if you want to get to, like, 300 plus, I think you can crack something in six months. And I think 18 months is a really good time to get, man, if you're starting from zero. And I think it's a great time to do that because, uh, the- there is a level playing field. There is a lot of AI PMs needed, and whether you were at Google before or not doesn't matter, or whether you were at Meta before or not doesn't matter because we need that talent. And if you are outside and have been iterating, solving ambiguous problems and building product with iterations and not by research, then I think this talent, the world is looking for. And if it's not happening, just hold on to it. By the way, it took me, like AWS told me no, like when Meta took me. Uh, Google also said no. Uh, [chuckles] so, uh, just a secret tip for you that it doesn't always work and then you just get double salaries and all.

    11. AGAakash Gupta1:15:22

        [chuckles]

    12. MYMahesh Yadav1:15:23

        It takes time. But if you're working on something that you're passionate about and you believe in the tech stack, and you are working as fast as, uh, any- anything you can afford to, uh, 18 months is a really good time for anybody to just get in and get to the top 1% of agentic AI PMs. And you can just join our community, listen to me every Friday, and if you're building what I'm telling you or using our learnings, I promise you, 18 months, I've seen nobody fail so far.

    13. AGAakash Gupta1:15:54

        What is that 18-month roadmap

16. 1:15:56 – 1:28:09

    18-Month FAANG Roadmap

    1. AGAakash Gupta1:15:56

       to becoming a PM at FAANG?

    2. MYMahesh Yadav1:15:59

       So let's divide it into your six milestones. Your first milestone is understanding and building your first good prototype. So what does that involve? That involves do you understand what goes behind the model, this intelligence piece in this diagram? So this intelligence piece is your LLMs. You can call this GPT-5 or this. How this piece is built, then how you connect it with your own knowledge of the company, your own databases. So this is how... So this is the thing that we all have, intelligence, and this is coming in machines these days. This is how the world works. Knowledge that is not in the model, which is, hey, what's the vocations in your company? What you like as a-- where you have visited in last one year if you're building a travel agent. Then your memory or signals, then your tools which can change the state of the world. So this is how the world works, current state of the world, and then your tools, which is how can you change the state of the world? And then guardrails, which is what you should do and what you can't do, or putting all the checks and laws or even checking by feedback, did we achieve the goal or not? And then bringing this all together to build a real working agent, that should be your goal in first three months. That you understand these concepts in and out, and you have built your first agent with all of these ingredients in it. That should be your goal for the first three months. After that, what you're looking for is either starting your own thing on side, which is a real good problem you have, which can be any problem under the sun. I don't care. Do you have PhD level expertise in that thing? Does it have unstructured data and complex decision-making? And if it has all these three, just start solving that problem. That's your next three months, and you should have at least 10 to 20 free users on that. That will teach you how to evaluate these things. So next, what we are learning is first three months you learn concepts, then you go and say build and evaluate. Once you reach that, that's your six months, that's your first iteration. And after that next six months, I really want you to get it into production. Whether you are hiring your own team of five engineers, maybe somewhere in India or somewhere it doesn't cost that much, and you are actually making it production and going from 10 customers to 100 customers. Another idea there is work for a startup for free. People will love to take that, and everybody wants people who have this much done and help them give them free hands. So that's your next six months. Build and give people the right thing. This will teach you how actually this build and evaluate iterations will work and bring the ambiguity and scale factor on your world. Once you have that, you have hit an year, then you start actually doing lot of... After this, you're trying to decouple any companies that you want to target their products and seeing what feature you will build. And you start working in their open communities. You run evaluations for them free. By this time, you should have your evaluation frameworks, and you are running evaluations and telling them what are their gaps.And you're doing it for top 10 companies that you want to work for, for their top 10 products, which everybody cares about. Whether it is OpenAI, Anthropic, Meta, Llama, start doing that. I've done that, and it does wonders. And then you create a small community around it, and you are building them, and you're saying, "Hey, you know what? I got these three engineers, I came up with this feature list, and I have checked it into your code base." Like at that point would like, okay, Meta won't hire you, but Google will hire you, right? Uh, nobody can stop you. Next iteration, you're even doing better than their PMs. And the beautiful part is all these companies have open sourced their model or some version of their model. Even GPT has an open source model now, and you can take that and show them how GPT-5 is so bad at instruction following, and take the open model and fix it, and do a side-by-side eval and publish it as a paper. And you can do that once you reach this six... one-year timeframe. After that, our students, like people who have done, have been our community, have done the courses, have been done all the paths. They get jobs before this step, but if they don't get it, I know a guy who just hacked the whole website and showed them their prompt and said, "Hey, maybe you want to do these checks in, otherwise people will see your prompts." And he got his first gig like that, and then he worked in... Now he works for Carfax building their AI stack. So it's not like people are desperate for you, but you need to be desperate for learning, and that's your 18 months, month on month, and every three months you should be hacking after this. And I don't... Means if you reach that stage, just ping me.

    3. AGAakash Gupta1:21:27

       [chuckles] This was the clearest, most actionable roadmap I have gotten yet, and I've asked quite a few people this question, so make sure you guys are taking notes on that. I wanna talk about this final angle of AI agents, which is if you're just a regular PM, not an AI PM, what AI agents should you be building to improve your efficiency on the job?

    4. MYMahesh Yadav1:21:49

       It's a really good question. I think one AI agent which all of us need is the guy who looks at our customer interactions. So we, uh, we live in a multifaceted or multimodal world, and we go to a lot of calls where our sales or our friends or things talk, talk about this. So always have a, a single agent which goes through every channel, every interaction where you can get feedback on your product, processes it and does something. I think that's a missing agent. I have not seen any tool. I've seen like some tools here and there. I think you should build it and iterate on it. So that's my number one suggestion. My number two suggestion is build something that you can run A/B tests at scale, which is this idea that before what you want to run, there's never enough people to run A/B test. And to be honest, when we ran them, it was such a, uh, you know, insignificant sample to make any judgment calls. But we live in AI now, and you can create thousand of different personas with small, small variations, put them into groups and show them your product, and maybe run a quick simulations of A/B testing or just simulation of who's gonna use their product most or love the product most. So one is this simulation should teach you how your product fails and it, for what customers. So that's one thing you should get out of that agent. And second thing, the output of that agent is that it tells you which personas will love your product, which, uh, which by the way, will be the people who have finished the flow faster, have stopped at the re-key points and clicked the feedback. So these two, and I think all of us should build a reviewer. I think people lose lot of, especially people who are new in AI, lose lot of their, uh, uh, lot of their brand in the beginning because they have these documents which are immature. They had used terms loosely, and not to their fault, it, they are just new. So I will build a reviewer. I think people who are building like, you know, PRD tools or create PRD tools, I think they're really foolish. They don't understand like how much is about credibility of the person and their intuition is, goes in that PRD. And if AI generates it and you think that can solve or that's how the organization will make decisions, I think we're not there yet. So I don't need that. What I need is I will write the PRD and I will check it with AI, but I want to have a reviewer who has looked at my manager, 20 reviews, how he commented, and can automatically comment on my PRD in his style, what he cares about, what he doesn't care about. And if I build that reviewer, everybody should be able to use it. Just upload whoever you are reviewing with, take their previous 10 reviews, upload them, upload your document, click a button, you got the first review. You address the things, and then you show up. I think that people will use and will be a great thing. So A/B tool for testing or a scale to run simulations of your customers and get insights, uh, which is based on this thing which actually watches different channels and give you insights. And third is a reviewer tool. I think these are the dream tools which I want to build or use, or I think everybody should build and use. And we are working on a couple of them if you want to help. Happy to take that help too.

    5. AGAakash Gupta1:25:26

       Love it. So we talked about how well these roles are paying, yet you've chosen not to continue on the path of these roles. I have to ask, this is the hot question now. You're on the hot seat. How big is the business of Mahesh?

    6. MYMahesh Yadav1:25:42

       Ah, I think that's a good question. Uh, so I think I'm very comfortable with the, the community and things that build. I won't give you my exact numbers, but you can look at how much our course sells for. We have grown 40% cohort by cohort. We started with 40 people, and we have grown 40% each cohort. This, we are on our eighth cohort, and we just served-- we are serving eight- 180 or so people. And we have divided into two because I never wanted to be, have more than like 70, 80 students in one group because I really wanted to have personal interactions and know them by name. So I run two classes now. So that's how we are making money. Uh, but I'm also working on building my own thing, and the money that we make here goes into that.

    7. AGAakash Gupta1:26:28

       Okay.

    8. MYMahesh Yadav1:26:28

       So not a-- like, business of Mahesh is a lot of transactions. I'm a medium these days, so it comes from here, goes here. And the goal is to bring the best AI capabilities in the hands of people which are least likely to use them without help. And with that goal here, we are trying to bring a lot of people who can't get to AI faster. And then on the other side, we are building tools for people who can bring AI to a lot of other fields, not just tech, so that more people can come in. And if I can be a catalyst in that reaction, that's business of Mahesh. And now you did-

    9. AGAakash Gupta1:27:08

       Love it

    10. MYMahesh Yadav1:27:08

        ... some math maybe already. [chuckles]

    11. AGAakash Gupta1:27:10

        I'm doing the math in my head. I'm trying. Okay. This was amazing, Mahesh. This was a masterclass in AI agents for PMs. Thank you so, so much.

    12. MYMahesh Yadav1:27:20

        Thank you for the chance. Uh, I always wanted to be here. I always looked up to you for, from the beginning, even when AI was not cool. I think growth was the thing. Growth was, I think, I-- it was an eye-opener first time for me. It means nobody talked like that about growth. It was more about building, especially in the companies I worked for. But you brought, and I think couple of people at that time brought this idea to life, which is it's an essential skill of a PM to grow the business, and I think nobody ha-else has contributed more than you on that channel. And with AI, you are pivoting, and you are iterating so fast that I can't catch up. So thank you for having me. It's just, I'm so humbled for giving me this experience. Thanks, Aakash.

    13. AGAakash Gupta1:28:04

        I'm the humbled one. Thank you so much. Everyone, check out his course

17. 1:28:09 – 1:28:52

    Outro

    1. AGAakash Gupta1:28:09

       on Maven, support his startup when that comes out, and we'll see you in the next episode. So if you wanna learn more about how to shift to this way of working, check out our full conversation on Apple or Spotify Podcasts. And if you want the actual documents that we showed, the tools and frameworks and public links, be sure to check out my newsletter post with all of the details. Finally, thank you so much for watching. It would really mean a lot if you could make sure you are subscribed on YouTube, following on Apple or Spotify Podcasts, and leave us a review on those platforms. That really helps grow the podcast and support our work so that we can do bigger and better productions. I'll see you in the next one.

Episode duration: 1:29:02