Six agents, one cron, a newsletter nobody had to write.
A daily learning platform for AI, DevOps and Cloud engineers. Six agents run on a cron, turn the morning’s firehose into something worth reading, and publish a newsletter and quizzes nobody had to write.
RSS feeds and Hacker News go in. Keyword and quality scoring drops what is off-topic. A Groq LLM writes a two-or-three sentence technical summary of what survives. all-mpnet-base-v2 turns each into a 768-dimension vector locally and for free. Supabase stores them behind an IVFFlat index, keyed on the original URL so re-runs upsert rather than duplicate. Then the day’s newsletter is built and up to five quizzes are generated from it. FastAPI on Vercel Python serves the API; React and Vite serve the reader.
sentence-transformers runs inside the Actions job, so embedding costs nothing and adds no rate limit to the critical path. The trade is that the embedding dimension becomes load-bearing infrastructure: all-mpnet-base-v2 emits 768 dimensions and the articles.embedding column must match it, so changing model is a migration and a re-embed of every row, not a config change.
The whole pipeline is one scheduled GitHub Actions run with workflow_dispatch enabled, so it can also be triggered by hand. For something that needs to happen once a day, a queue would be infrastructure to operate in exchange for latency nobody is waiting on.
Article text comes from the RSS excerpt at crawl time and is dropped after summarisation. It keeps the database small and avoids storing other people’s content — at the cost that repairing a bad summary means re-fetching the original URL, which is what backfill_summaries.py exists to do.
Groq retired llama-3.1-8b-instant on 2026-08-16. Every summarise call began returning 404, and a per-article except fell back to summary = title. The pipeline embedded those titles, persisted them, and exited 0. Six consecutive green runs published nothing but headlines, and it cost thirteen days to notice — while a second model, llama3-8b-8192 in the quiz generator, had been dead for roughly a year. summarize_articles now raises when more than half the articles fail, so the run goes red before anything is embedded or written. A per-article fallback is fine; a wholesale one is a failure and must not be allowed to look like success.