LLM-curated content feeds beat platform algorithms
Build a niche curation tool that uses an LLM to filter a large public text corpus (Hacker News, Reddit, forums, niche subreddits) into ranked or recommended video/content lists for a specific community. Package it as a micro-SaaS or browser extension that beats generic platform recommendations for a defined niche audience (developers, makers, researchers) and sell subscriptions or a one-time curated feed.
What to build
A niche video-curation web app for one tech community (e.g. indie hackers or AI builders) that uses an LLM to rank and filter video links surfaced from a large public text corpus like Hacker News archives, Reddit, or niche forums, refreshed on a schedule.
Karpathy showed that an LLM plus a large historical text corpus (20 years of Hacker News) can be turned into a working curated content platform in a weekend, proving niche LLM-curation tools are now a solo-buildable category rather than a team project.
Demand
Builders and researchers in fast-moving niches are tired of generic platform algorithms burying relevant content; a credible, well-known builder just demonstrated the exact pattern works, which will push other solo developers to copy it quickly.
- Hacker News / tech signal reportNews
Report that Andrej Karpathy used Claude to build a curated video platform over a single weekend, leveraging 20 years of historical Hacker News data as the underlying corpus for curation.
- Signal metadataAnalysis
Opportunity scoring (0.68) identifying a replicable pattern: micro-SaaS or browser extension products that use an LLM to filter large public text corpora into niche-specific ranked content feeds, monetized via subscription or one-time curated feeds.
Stack
- Claude API
- Hacker News Algolia API
- Reddit API (PRAW)
- Postgres
- Vercel
- Supabase
Solo + AI difficulty
Easy: pulling HN/Reddit data and calling an LLM to rank it is a weekend task with AI coding help. Hard: picking a niche narrow enough to beat generic feeds, handling API rate limits, and building a UI people return to. MVP in about 3-7 days.
- Entry threshold
- A solo builder can replicate this in days using an LLM API (like Claude) plus existing public datasets or APIs (HN API, YouTube API, Reddit API); main cost is API usage and some engineering to scale filtering across millions of comments, no licences needed.
- Window
- 3-6 months
Where to find first users
- Hacker News Show HN post
- Product Hunt launch
- Indie Hackers community
- Niche subreddits for the target audience
Counter-signals & risks
A single high-profile builder (Karpathy) with deep AI/ML expertise and reputation may not represent a replicable path for typical indie developers or small teams attempting similar niche-curation products
Reliance on third-party public corpora (Hacker News, Reddit, forums) exposes the product to API access changes, rate limits, scraping restrictions, or terms-of-service enforcement that could disrupt the business model
Large platforms (YouTube, Reddit, Google) could integrate similar LLM-based niche curation directly into their own products, eliminating the differentiation advantage of a standalone micro-SaaS tool
Original title: Andrej Karpathy built a curated video platform in a weekend using Claude and 20 years of Hacker News
Related signals
- AI agents#2
AI Hallucination Detector for False Execution Claims
Build a lightweight AI transparency and reliability monitor — a browser extension or web app that detects and flags when an AI assistant (ChatGPT, Copilot, Gemini, etc.) claims to have performed an action it structurally cannot perform (listening to audio, browsing, running code, etc.). Target paying AI power users, journalists, researchers and enterprise teams who need an accountability layer on top of LLM chat. Monetise via a small subscription or one-time purchase.
Demand7/10est.Buildability7/10est.Competition10/10est.via OpenAI Developer Community0No named competitors - AI agents#3
Central memory layer SaaS for multi-agent AI workflows
Build a lightweight SaaS or self-hosted tool that acts as a central memory and context layer for AI agents via the Model Context Protocol (MCP), letting solo developers and small teams sync shared context, persistent memory, and state across multiple AI tools, agents, and automations. Target vibe-coders, indie hackers, and small dev teams who are stitching together multi-agent workflows and losing context between tools. Offer a simple dashboard to manage, version, and inspect memory, plus an MCP server endpoint they can drop into any compatible agent setup.
Demand7/10est.Buildability7/10est.Competition6/10est.via Product Hunt33 competitors - AI agents#3
Personalized AI advice layer beyond generic responses
Build a personalization layer or AI advice differentiation tool that learns a user's specific context, constraints, and goals before generating recommendations — so the output is genuinely tailored rather than generic. Target knowledge workers, freelancers, and coaches who rely on AI assistants daily but are frustrated by cookie-cutter responses. Could be a Chrome extension, a prompt-engineering wrapper, or a lightweight SaaS that stores user profiles and injects rich context into every AI query.
Demand4/10measuredBuildability9/10est.Competition6/10est.via Hacker News33 competitors