ai-news

Keeping up with AI had turned into a part-time job: two or three hours a day across Twitter, HackerNews, arXiv and half a dozen newsletters. I wanted the summary, not the ritual. So I built it.

ai-news gathers what gets published every day across a dozen sources and leaves me a briefing I read in ten minutes. Underneath, almost all the work is mechanical and boring on purpose: deduplicate, score by real interest, order things so a single big story doesn’t swallow the feed. Rules, not magic.

The language model comes in at the end, and only where it’s genuinely needed. A first keyword filter —free and instant— handles most of the 2,200 items that arrive each day and trims them to about 150. The LLM only classifies the doubtful cases and writes the one-line summary. It’s the pipeline’s last resort, not its engine.

I didn’t design it just for humans: a public MCP server lets other AI agents ask it the same questions I do from the web.

One detail I like: the code was written by agents, almost line by line. The decisions that matter —what to discard, where to put the model and where not to— were mine. That division of labor is, at bottom, what all of this is about.

---

How it’s built — FastAPI · PostgreSQL + pgvector · React · MCP

Sources: HackerNews, arXiv, Reddit, RSS, GitHub, HuggingFace

Try it at pguerrero.me/ai-news — or let your agent query it at pguerrero.me/mcp