Inspired by Karpathy, I built an AI-curated video hub for the startup community
A few evenings of watching random stuff on YouTube, and suddenly my professional recommendations are gone. I watch a lot of startup content: founder interviews, VC discussions, growth strategies, and tech deep dives. But I also use YouTube for entertainment. And after a few evenings of guilty pleasures, my homepage becomes a completely different place.
My professional interests haven't changed. My recommendations have. I could create a separate YouTube account, of course. But I wanted something more useful: a dedicated space for discovering professional content, organized around my interests rather than my recent viewing habits.
The idea came from Andrej Karpathy
On November 1, I saw Karpathy's HackerTube experiment.
He used Claude to process 43 million Hacker News comments and extract 6,454 recommended YouTube videos.
Some people mocked the idea of "building YouTube in two hours." But I think they missed the interesting part.
You don't need to rebuild a massive platform to create a different experience on top of it.
YouTube already handles hosting, streaming, and the enormous content library. What if we could build our own discovery and recommendation systems on top?
What I built
I built TheBold Hub — a personalized video discovery platform for founders, VCs, and the startup community.
Unlike HackerTube, which uses historical recommendations from Hacker News, TheBold Hub continuously discovers new content from a curated collection of YouTube channels.
Here's what it does:
Curated sources: Around 200 YouTube channels covering startups, venture capital, growth, AI, and technology.
Automatic discovery: New videos are collected through the YouTube API.
AI filtering: A lightweight Perplexity model checks whether each video is relevant.
AI categorization: Relevant videos are automatically organized by topic.
Personalization: Users can customize their feeds around their professional interests.
Community contributions: Members can suggest additional YouTube channels.
Discussions: Founders, investors, and tech enthusiasts can discuss videos without leaving the community.
The videos still play through YouTube's embedded player. I'm not hosting or duplicating the content.
The technical approach
I built the application using Claude Pro, with Claude handling most of the implementation.
The basic pipeline looks like this: Curated YouTube channels → YouTube API → AI relevance filtering → AI categorization → database → personalized feed
One interesting discovery was how inexpensive the classification step could be.
Using a lightweight Perplexity model, my initial estimate for filtering and categorizing 1,000 videos is approximately $0.01.
That's not the total cost of operating the platform, of course. There's hosting, maintenance, and other infrastructure. But it makes automated curation across hundreds of sources economically practical.
You could apply the same approach to educational videos, scientific lectures, design tutorials, or almost any other niche.
💬 I'd love to hear your thoughts, ideas for improvement, and recommendations for YouTube channels I should definitely track!
Replies
Entertainment eating your professional recs or otherwise is so real. Cool product! Question about the personalized feeds: do I pick topics once, or can I tune them as my interests shift?
@vehsakam Thanks! All 8 topics and 30 categories are always there to browse. On top of that, the feed learns from what you do: opening a video, watching past 25/50/90%, saving and commenting all count, with deeper actions weighing more. Leaving within 30 seconds or hitting “Not interested” counts against a topic.
@vehsakam Sign in, and after 5 signals, you get a For You section, topics you care about move up, and one spot always stays outside your interests.
@juliayugo Love the detail about one spot always staying outside my interests. That's a thoughtful way to fight the bubble indeed. Nice work, Julia!
@vehsakam 💜
Right now, AI pulls in and filters videos from 200 English-language sources, organizing them into a predefined set of categories. Anything that doesn't fit gets filtered out.
For example, you won't see a podcast where Elon Musk talks about the challenges of fatherhood. But an interview with Ivanka Trump discussing AI or venture capital? That will definitely make the cut 🙃
If you sign in, you'll also get personalization: your homepage feed will gradually adapt to what you watch.
I watch YouTube for work while walking between my home and my study, so my feed has exactly this problem. I had a look, and the top video when I opened it was YC's talk on why the harness matters more than the model, which is the kind of thing I'd actually watch.