This video features an interview with Zachary Huang, a PhD student at Columbia University, about his open-source AI tool, the AI Codebase Knowledge Builder. The tool uses AI to analyze codebases and generate step-by-step tutorials, simplifying complex code for developers. The video also explores PocketFlow, a 100-line AI framework enabling "agentic coding" where AI agents build other AI agents.
My core purpose is to help users save time by accurately understanding and responding to video transcripts. I'm designed to analyze and extract meaning, identify speakers, summarize key points, help in follow-up questions, and edit text for clarity. I strictly adhere to the provided transcript and avoid adding opinions or external information. While I can perform other tasks like creative text generation or translation, these are not my primary functions. COFYT may make mistakes, so double-check its responses.
I cannot directly create a downloadable text file. I am a language model; I operate within this text-based interface. However, I can provide you with the full transcript text without timestamps, which you can then copy and paste into a text file and save yourself.
What if you could have an AI that could explain any codebase to you in just 5 minutes and what if you could turn any GitHub repo into simple easy to follow step-by-step tutorials well today you're in for a very special treat we have Zachary Hong who built just such a tool and it's an open-source tool that anybody can download and use right away this tool enables an AI to read through entire code bases and then transform complicated code into stepbystep easy to follow tutorials whether you're an engineer that inherited some messy code or you're a vibe coder who wants to build cool projects with AI this is a godsend gift when I reached out to Zach I wanted to learn more about this tool but actually I realized that he was building something more grand than just this super tool he was building an AI framework with just 100 lines of code and this simple yet powerful framework will enable agentic coding meaning now AI agents can build AI agents isn't that super cool now let's hear it from Zach hi Zach it's so great to meet you can you give a quick introduction about yourself yeah thank you for having me here i'm Zach i'm a PhD student at Columb University i've been working on database systems for over four years but over the past two years I've been working on large range model systems and I'm about to graduate and join Microsoft research in a bit over a month again thank you very much i'm very glad and happy to chat about my recent works so now I came across you because of this AI codebase knowledge builder and I'm sure the people who are watching this right now would love to learn more so can you kind of tell us more about it how does one use it and is it something that I could just give you a GitHub link and have you generate this knowledge builder for me yeah so the idea here is whenever you are joining a new team or whenever you are looking at a new open source projects it's pretty overwhelming to just read through this code bases and understanding figure out what's going on under the hood so we just use AI to help you make a pass and help you generate such tutorial so we have already applied this tutorial across multiple popular GitHub repository we can just give you some example which one do you are you interested in just looking at so Zach is kind enough to let me choose any repos so I shared one which is a YouTube summarizer cuz I like to learn a lot from YouTube and sometimes I just want to summarize the YouTube videos so I can have the notes so I just shared a link with Zach yeah in order to run this new repository what you're going to do here is you're going to just follow the get started instruction here so we're going to clone this repo which you already done and we're going to install the dependency which I've done in the past we're going to set up the larger model course so essentially in this models you're going to set up by providing your own project or if you can or if you are using this AI studio you need to provide your API keys but if you are using cloud 2.7 or open a01 you just implement your own function that takes the string input and call the logic model return the response but after you set everything up essentially you just call this python function here so here let's just copy paste this new repository URL to the page let's give it a name maybe let's call it YouTube summarizer and it will run it will so what it's is currently doing here is it is currently crowing the all the files from the GitHub repository and then it is calling the live model to understand hey what's the most important concepts inside of this repository and how are we going to present this concepts and easy to read way for the audience can I ask a question so you know how there's like a limit in context window so why is this tool that you're building able to digest the whole code base ignoring that that limit well the limit of the current large range model is actually pretty large it's let's take Gemini 2.5 Pro as example it currently has 1 million tokens it's more than enough to understand most of the codebase it's not really the contest limit issue here but I think the issue is mostly like how do you manage the contest because you can just dump a lot of options into the 1 million token but there is a phenomenon called loss in the middle which means that the model will just neglect the middle part and will focus only on the start and end part of the contest which is pretty human like when we human listen to these you know messages reading to post we just look at the beginning and ending regarding the part uh so what do we do here is we have this a workflow so let me just show you the the design of the of products so we start by identifying the high level abstraction relations concepts we're going to teach the users so this part we just take the whole the whole code into the context but then we're going to write a chapter one by one so here we already starting to write a chapter one for the front end application structure for this individual chapter writing it will only focus on the kind of the files instead of the repository that matters for these concepts so it doesn't take the whole repository it only takes those relevant and which one are relevant is alo decided by large model on the flies so you can keep focus on the most relevant part and the best quality you know chapters uh for the tutorial i see and you being like a system person you mentioned does your background make this easier for you than other people because this is like thinking about systems right yeah it's yeah it's like how you design a system when you have like a very large workloads previously maybe it's a compute or maybe a data crawling all kinds of these different systems you're going to identify what is the key bon here what is the most complex part of the task here and how do we decompose how do we build a microservices that talks with each others each handling individual task here so I do think like this system designs or this mental model helps a lot in terms of designing and large model systems in this AI era and how does somebody gain this kind of mental models is there something that you would recommend for people who are interested in developing a better system mindset or do they just choose the AI and learn yeah I would say it's first of all the whole large model applications systems is pretty new concepts and I I'm still in the stage of explorations i'm still trying build different systems get my intuitions uh what's the trade-off what's the results from different designs so it's pretty early age there's no good tutorials for that and I'm trying to make tutorials based on what I have already learned but still a nent area but I do think like traditional system designs is helpful in somehow like how do you design a service how do you build a application that help a bit I guess and also people can follow you on YouTube to learn more about your thinking and mental models yeah I'm very new to YouTube i just made YouTube over the past few weeks so I think I have a decent technical skill but my presentation is not that I'm a follower thank you thank I'm still learning i'm still trying to practice my skills in terms of YouTube so would greatly appreciate your support and criticisms yeah people who are watching follow him follow Zach thank you thank you yeah the large model call is a bit slow here but that's because we're using the best models we are trying to ask it to do a pretty complex task here actually one more question so the default here that cuz I know that it's like writing chapter 1 2 3 4 5 and the default model that you're recommending people to use is Gemini 2.5 well I it it changes right gemini 2.5 is the best model last week uh but so what is it this week yeah OAI just announced 03 04 mini i'm not sure how good they are i haven't tested them out but maybe they are better i don't know but just as of last week they are the best model i see so is it done or not done it is done let's check out the results here so we have the YouTube summarization here and we have the the results it says this project is a web application that use artificial intelligence to create a summary of YouTube videos user paste a link and choose options like the model the app fetches the videos transcript and use AI to gen generate a summary that shows the progress in real time and it even generates like a diagram as a system person I very much love this kind of diagram because text is sometimes just pretty linear but you really want to have this two dimensional understanding of what's going on and how this each components rel to each other so it's something I really emphasized when I was designing this actually people who are interested in the system thinking should use this tool because with this graph they're learning systems thinking right yeah yeah and yeah it's like I've encoded a lot of my way of how I write this tutorial instead of this prompt so I think the it so here I provide a different rules on what you should do when you're generating these chapters you should you should begin with a high level motivation if the actress is too complex break down to key concepts the code block you should also make it minimal make it simplified you should also understand you know you should also illustrate the complex concept using this mermaid diagram so been code of instead of the prompts and the model just pick it up and help generate this kind of the a very nicely visualized architecture for us using mermaid diagrams wow that is so cool okay can I keep looking at this tutorial that's generated yes so we have different chapters so each box here correspond to a different chapters we start from the high level front end application structures so this is a TypeScript issue i'm not a really friend person so I don't really know what's going on here but it breaks it down into the application the layouts the pages the the roots and also provide another nice sequence diagram on what's going on when people asking for YouTube samurai and how does this different call different files instead of this application work together wow this is so cool it's such a good learning resource it's almost like how computer works And but this is like how this repo works and then it's breaking it down step by step exactly and like this is only chapter one here you also have this chapter two more on the UI sides chapter three how do you get this transcript from the YouTube audio chapter four the how does the AI comes into play and work with this transcript to generate a summary and you have the whole summarization pipeline for you so it's starting from the most userfaced section on the UI parts and then a step by step dig deeper and deeper into how does it work internally what's the workflow what's everything what's the back end pipelines working under the hood so dig deeper but also in a very progressive way organized way yeah also one question this is ran in cursor so theoretically you can open the right hand side panel and chat with the agent on this right if I have a question about a class and I wanted more detail about it is it possible for me to ask about some concepts in here yeah we can just ask cursor which also is Gemini 2.5 Pro it is reading the file so it's not just purely chbt actually contextualize your answers based on these tutorials and they're talking about how and how CSS work how shenan works and what's the difference here oh my god Zach this needs to be productized this is so valuable to people whoever is watching this are so lucky because now they know they understand another way to learn and this is so cool i don't know if people are recognizing the value in this tool right here yeah if you have a new codebase you can first ask it to generate a tutorial then step side by side you have a teacher along you if you have any question in the middle you can just ask through a chatbot and it will context your questions in the context of this tutorials it can even do some web search it can help you analyze and generate such a great answer you can even customize this tutorial generation workflow based on your own needs yeah you can for instance if you have your own knowledge if you have your own writing styles you can even add more images to this tutorial based on your demands and this is open source so people can either contribute to it or they can just like download it and clone it actually what you did is just you just use one command and pasted the GitHub repo link in there and voila this is there yeah it's very simple it's fully open sourced we provide example GitHub repository common from autogen browser use crew AIDS py numpy requests any kind of the popular repository you can use this tool to generate a very friendly tutorials and you can just clone this repos install and set up your own large models and you can just paste your any repository you want to learn about just run this a single lines of the command and it will generates a tutorial for you in around five minutes it's fully open source and actually we just get a lot of different poll requests over the past few weeks people asking for multi- language support people asking to making the call a model friendly by integrating and providing sample of different models adding virtual environment support local G repository and so on and uh you know that's so cool it's free i can't believe it yeah and I will appreciate your contribution for any of the new features you want and just send a post to our review and just send a new PR and I will review and merges oh my goodness this is such a treat this is crazy i'm so happy to learn about this and this product i'm pretty sure I'm going to use it a lot because I just a lot of times I'm vibe coding and sometimes I have no idea what this thing is doing and I just could use some help and I think this is like the perfect t like tutorial or tutor companion that I could use as a vibe coder yeah exactly i think even for vibe coding a lot of time it's like you want to get a control at a high level even if you are trying to have this like a self-driving experience you still want to get a control you want to know like what the how does the root works what does high level how you want to do this kind of high level planning here you don't want to just like the card drive like headlessly so I think like even for different code base even for this web coding project you still want to understand at high level what's the architecture what's different components And why do they make sense here so you can you can criticize on different components or you can know opt out optimize different parts separately this kind of a control will help you make your vibe code projects more reliable more maintainable for long-term use oh wow okay digging deeper into this because I know that you actually are a really good writer and you publish some Substack articles right and yes I have been reading them and learning more about this and you also mentioned that pocket flow which is this AI framework is what what you use to build this right can you tell us a little bit more about that because it's a 100 lines of code but it's super powerful it's an AI framework so can you tell us about what Pocket Flow is what's an AI framework and why is this 100 lines of code so powerful and enabling a gentic coding yes pocket flow is a large model framework in just 100 lines as you mentioned here think of pocket flow maybe similar to n or zapier but is a bit more technical so based on this MB nap here you can build all kinds of business workflows to automate maybe your linking maybe your Slack or Discord communications but with Pocketflow you have a similar abstraction but you can build all kinds of great collage model applications like the tutorial codebased knowledge builders that we just demonstrated a couple of minutes ago and I really like the analogy that you made for me to understand pocket flow better which is the kitchen analogy can you explain the kitchen analogy to the people who are watching this yeah the way to think of pocket flow here is it's a graph but you can think of it as like a kitchen in a kitchen you have different stations it could be for chopping for could for cooking could be for plating so that's what each node is for and you have this flow essentially connects this stations so the flow decides which station to go next from the previous one so for instance after you have done with the chopping you may want to go to the cooking after you're done with cooking you may want to go to plating so flow decides which station or which note to go next and then you have a share store it's well like all these different stations stores these ingredients every stations can use this counterpart to store these items and can grab the ingredients for them to use yeah that's the high level idea of the of the kitchen analogy of the pocket floor so essentially you want to decide what each station does you want to orchestrate them you want to connect them using the flow and finally you want to put all this data all these ingredients into a contot that every station or every nodes can get access to and public flow based on these graph abstractions can express all kinds of the popular descent patterns be it workflow chat agent channel of dots retrieval augmented generations even multi- aent supervisors and so on and we provide the tutorial of cookbooks for all kinds of different design patterns you can check this out we also building different more complex model applications include in including these codebased language builders for you to understand how can you build cool applications on top of pocket flow so this is a framework for people to build more AI applications more easily and just 100 lines i think I read the article that you wrote which is that simplicity is king but can you explain to us why less is more in this scenario i guess that really matters a lot in the age of large model or in the age of maybe V coding because most of the time we we don't write code lab we're not that in the age of cursor wind surf and all kinds of coding assistant we just delegate this low level coding for this language models and I think the characteristics of this coding agent is very different from human so for this coding agent s they are capable of understanding and coding at a pretty low level uh so we want to have this a clean interface for them to code against and they can build everything based on these building blocks here and the problem compared to other framework let's say lchan or llama index is that they are pretty complex they have like tens of hundreds of thousands of codes and they and and the issue they call is like the lanterns abstraction is overly complex so what I find personally that is very helpful when it comes to let AI building application for me here is that they don't need a complex abstraction what they need is a very clean very easy to understand small interface to program against and that's what pocketflow provides it provides a very clean very straightforward interface of a graph and based on this graph interface it can build all kinds of cool application for you so I've also created tutorials on how can you use pocket flow to ask GPT assistant uh charge GBD and cloud projects or cursor winds of client to help you do this application development essentially you want to pass in both the pocket flows source codes as well as pocket flow's documentation uh how to use pocket flow with the steps to design different design patterns and it will pick up and build for you immediately well and that's how you build the AI codebased knowledge generator right exactly so the kind of the steps I work with cursor is that I start by this system design documentation now the design is pretty important it provides you like a high level control over your codebase and even for this design to be honest I don't write it word by word i still ask curs to help me make a pass but I'm very critical i'm I will criticize the design i will ask cursor hey improve this part this node is not well designed this part is too complex and cursor will iteratively improve it improve the designs based on the guidance and to a point where the design is clear and is easy and I'm satisfied then I ask cursor to write the whole codes for me basically implements these nodes the flow the graph abstraction and all these codes you see here they're not written by me it's just written by cursor uh as long as you have a good clear design here all these codes can just be truly implemented because current model is pretty good at this low-level implementation wow and I saw that you have a lot of like projects that you've built using pocket flow and they're all open source people can access it can you show us some of the other projects that you've done i guess another cool project is just build cursor with cursor the idea here is the cursor have cursor is very helpful in the sense that it doesn't only chat with you it can help you edit the files it can look up your codebase and so on but under the hood it's still like agents so what cursor what it does here is it has different file operations read edited files and we can design such graph like structure starting from user request we have a main node that decide what's the next step here and we have all these smaller nodes to perform different action for you but this is a this is not fundamentally different from the codebase generator it's it's a more complex graph it has this looping branching but still graph as long as you can provide a well designed documentations cursor can implement the whole flow for you automatically and you mentioned graph which to you said to think about it like N8N and Zap year or something like that i see that a lot of people are building and creating MCP servers and what's the relationship between that versus pocket flow mcp is more like a standardization it's like when you want to have different tools your agents to be used by others potentially in different processing in different machines in a different physical locations how would you communicate with each others previously we have different ad hoc way to communicate but MCB kind of standardize it by providing mostly like what's the request schema look like you can think of MCD server as a different ways to perform actions so maybe let's go back to this okay this graph so you see like in the main decision agents after it decides it can take different actions right from read file delete file search list directory and so on but you can replace all these nodes with kind of like a MCB server so MCP server provides the agents hey this is a list of the actions you can perform you can read so it's the nodes it's the nodes it's the nodes that present the options to the agents and after agents decide which action to take it will perform the actions on the behalf of the agents it's a note instead of the loop but just named as MCP sorry for the interruption but I bet you didn't know that I have a free school community that I would love for you to join i'm in the process of putting together some really cool AI resources and I would love for you to be a part of the community and meet like-minded people who are building cool stuff with AI so I hope to see you there thanks and how big is the team behind this pocket flow um well it's it's uh it's a side projects so a bit a little bit history about pocket flow it starts with my previous projects which is called cocoon my my PhD is about database i was working on this cocoon projects at Clint University for over a bit over one year during that time we made a different publications and this projects is it's also not big team but with a couple of mass students at Kban University i would say around two or three people but then pocket flow is really a very side branch out from cocoon we have this initial crappy version of pocket flow that have this nodes and graph concepts but branching out from cocoon and it's a sad project that I don't think I I initially don't anticipate that much of the the attention but it turns out to be very large and I was keeping building it i was just writing tutorial and working on that mostly over the past few months so it's a very early projects it's just my my mostly myself but thanks to the open source communities we receive a lot of contribution over the world wow people who are successful they seem like an overnight success but actually Zach has been working on it for a long time and he has been doing his PhD for what five six years and so he had all these systems background to put into this yeah thank you yeah I've working on cocoon for I would say around a bit over one years during that time I've been in different large model applications so it can just briefly shows you how do you build these data pipelines how do you do this retrieval augmentations over data how do you do this data cleaning data profiling and so on now this maybe sounds a bit boring because it's my my domain everything was databases so this kind my niche domain but since I'm almost finished my PhD and I have a more freedom to do to generalize these ideas so I was working on separating this workflow this graph abstraction outside of the the niche domains trying to make it a framework trying to apply to general programming coding development products so that's how it starts from this niche databases into more general projects but yeah I've been working on that for like over a year wow that's so cool i wanted to know though if people want to support you guys because you're giving all of these amazing tools and framework for free open sourcing it to everyone so how can people support you and your mission and what you're doing oh just come to my YouTube come to my substack and visit my repository contributing to do the open source project i have a lot of fun doing this all this building and the writings it's not really I would say it's really a leisure for me to doing all these works so I'm very happy i don't need any support beyond that but I will appreciate your engagement your com and your contribution especially for the open source projects and do you have a discord channel as well i would love to join yeah please join discord channel we're discussing most of a lot of people asking like hey how is this project works what's the recommendation for that what's the recommendation if I do want to build another projects people helping each others having these technical discussions please join this discord if you want to become part of the community okay thank you so much Zach for joining us today and I really learned a lot from you so hopefully we could stay in touch and now join your discord and maybe I'm going to try building with pocket flow and make my own AI agent build something cool that's what agentic um coding is right exactly exactly i think in the future like it's less and less about human coding human should do the design very well the human should have this high level understanding and after you figure out every high level science the model just get better at this low level implementations that's the kind of spirit of agent coding yeah and they can get better at this high level understanding by using your AI codebased knowledge builder and they get better at that you can follow a very positive dilute from this from different projects yeah okay thanks so much for joining us today it was so much fun thank you thank you for having me here bye and that's it for this video let me know what you think about AI codebase knowledge builder as well as Pocket Flow and be sure to check out Zach's YouTube channel and his Substack because he's just sharing such valuable content out there if you want to show support make sure to go to his PocketFlow GitHub and give him a star and if you have any questions ask in the comments below thank you so much for watching and I'll see you in the next video
