About
Code fetches and transcribes podcasts; an agent picks what matters and writes what was said. Code then checks the agent's work before anything is published.
How it works
Deterministic code reads feeds from dozens of investing, tech and earnings shows, gets each transcript from the cheapest good source (the publisher, YouTube captions, or local Whisper), and stores the result. An agent scores episodes for new, checkable executive or investor signal and writes claims, each with a speaker and a timestamp. Code validates the output against a schema and grounds every claim in the transcript: the episode exists, the timestamp falls inside it, numbers are spoken near that time, and no long passage is copied.
Daily runs pick three episodes. Weekly runs rank about forty and publish the top fifteen. The source and the full method are on GitHub.
Caveats
Claims are checked against transcripts, not against company filings. Speakers may be talking their book. Whisper transcripts can misspell names. Nothing here is investment advice.
Feeds for agents
- /runs.json: index of runs with date, kind, counts and paths.
- /claims.jsonl: every claim from every run, newest first, one JSON object per line.
/runs/<run>/claims.jsonl: one run's claims. Each claim hasspeaker,timestamp,tickersand alinkthat opens the source at that moment.