Podcast Transcript AI — transcription at scale

The product

Paste a podcast link from Apple Podcasts, Spotify or RSS and get a transcript, summary, chapters and FAQs. Transcripts are published in a public, searchable library, and developers can use the same engine through a paid API.

My role

Another developer built the first version. I took over in September 2025 and have written nearly all of the code since — about 576 of 593 backend commits — including the move to self-hosted transcription.

What I built

• Self-hosted transcription. Replaced a paid speech-to-text API with whisper.cpp running on our own GPU server, with three model sizes.
• A queue across two servers. Jobs live in MongoDB with paid, free and crawler tiers; workers on two machines claim jobs, recover from crashes and never run more than the GPU’s memory allows.
• Long audio. Files are split into chunks with their own time budgets, timestamps are stitched back together, and progress is real.
• Cost-controlled AI. Summaries route between a local model and a hosted one, with a daily spending cap — added after an outage in which an empty API balance published 531 transcripts without summaries.
• A library that grows on its own. A crawler reads the charts, finds each show’s feed and transcribes back catalogues when the GPU is idle, so paying users always go first.
• Finding the audio. When a platform blocks downloads, a fallback chain searches other public sources and matches the episode by its duration.
• More features. Speaker labels, search across every transcript, chat with an episode, PDF/SRT/VTT exports, Notion and Obsidian export, and a public API with credit plans.
• A second brand on the same backend. A sister transcription product runs on this backend with isolated accounts, sessions and payments.

Hard problems

• A clean-up step that froze the server. Removing Whisper’s repeated lines was quadratic and blocked the server for minutes on long episodes; I rewrote it to run in bounded time.
• Billing that’s right every time. API refunds are claimed exactly once, and an append-only ledger records every credit change.