Podcast Pipeline
Transcribe every podcast you follow on a daily cron and search all of it semantically. GPU transcription on Modal, Postgres on Neon, about $25 a month all-in.
My verdict: I use it to mine niche business podcasts for client research. Asking 'what has anyone said about pricing conversations' across hundreds of episodes changes how you prepare.
What it is
Two Modal apps and one Neon database, no servers to manage. An ingest app runs on a daily cron, pulls new episodes from your registered feeds, transcribes them with Whisper on a GPU, and embeds the chunks. A query app serves a permanent HTTPS endpoint for semantic search and full episode reads.
Who it is for
Anyone whose niche has good podcasts and no time to listen to them. Consultants mining industry shows for client language. Writers looking for the exact quote they half-remember. Me, feeding the summaries into my morning brief.
The economics
GPU transcription on Modal costs a fraction of the hosted APIs: roughly $0.60 a month for five hour-long episodes a week, versus about $7 through the OpenAI Whisper API. All-in with the database, plan on about $25 a month.
Setup outline
- Create a Modal account and a Neon Postgres project, set one shared secret.
- Deploy both apps (two commands). Modal prints your permanent query URL.
- Register feeds with a curl call; the daily cron handles everything after that. Backfill by date or episode number when you add a show with history.