Architecture of an autonomous startup-idea generator

Hi HN, OP here.

This project started on a whim a few months ago. I was curious if it was possible to automate ideation by feeding raw news into an LLM and getting coherent, viable business concepts out, rather than just generic summaries.

I ended up building a full pipeline (10 steps, db state driven) to test the theory. It runs daily on a DigitalOcean droplet, scans ~500 articles via EventRegistry, and uses Gemini 2.5 Pro/Flash to synthesize the trends and post a fleshed out idea to Ghost CMS each morning.

The stack is Python 3.13, Pydantic AI, and Postgres with pgvector. I found Pydantic AI particularly helpful for handling the structured outputs, and Logfire was a super helpful bonus. The vector search turned out to be critical for semantic deduplication so the agent doesn't get stuck in a loop suggesting the same ideas every day.

Happy to answer questions about the architecture or the prompt engineering logic!