MinusPodJev sits beside MinusPod and answers its ad-detection requests with Jev instead of a chat model. It speaks the OpenAI-style interface MinusPod already expects, so nothing in MinusPod has to change.
Jev scores one transcript segment at a time, and the proxy joins the flagged segments into spans, sorts them into categories, matches sponsors from MinusPod's list and returns the answer in the format MinusPod wants. There is also a native endpoint that takes segments and returns spans, and a status page with live settings.
The repository also holds an offline benchmark that measures Jev against chat models on real episodes, with the cached results and the commands to repeat it. It is a proof of concept, and chapter titles still need a second chat model.
What you get
- Full Python source code under the MIT license
- The proxy, its Docker setup and a status page
- The offline benchmark with its corpus, reports and reproduction commands
CodeRabbit participant
Jev
Python
Discord
FastAPI
React
TypeScript
