Overtone
Analytics for podcast producers. Upload a finished episode and it tells you what was
actually said: word and term frequency, how often each recurring bit came up, and a timestamp
behind every number so you can go and hear it. Search a whole back catalogue for the moment a
subject came up, and name the running bits you want counted across every episode. A two-hour
episode gets split up in the browser and uploaded straight to object storage, because a
serverless function only accepts about 4.5 MB of request body, and transcription is paced
against a provider allowance that refills as it goes, so the work outlives any single request.
Behind that it's Clean Architecture across four layers with the dependency rule enforced in CI,
ports and adapters covered by contract suites that run against the fakes and the real
implementations, over 950 tests, and every decision written down as an ADR. Built for Brazilian
Portuguese first, with the language chosen per episode, and bilingual throughout. It runs on
free tiers with no payment method anywhere, which is also the catch: the hourly transcription
allowance is small and there's about a gigabyte of storage, so it's slow and can't hold much.
That's the tiers rather than the code, and paid infrastructure scales the same app without a
rewrite.
- Next.js
- TypeScript
- Tailwind
- Drizzle ORM
- Supabase
- Groq Whisper
