// the take
A macOS app that listens in on a live interview call, transcribes audio locally with whisper.cpp, and pipes the transcript to GPT-4 to generate real-time answers and code analysis. It's built for someone doing mock (or actual) technical interviews who wants an AI prompter running in the background, paired with a Firefox extension that reads the live coding editor.
Whisper runs locally via whisper.cpp instead of shipping audio to a cloud STT API, so only text (not raw audio) ever leaves the machine. The BlackHole multi-output loopback setup is a genuinely clever trick for capturing both sides of a video call without feeding the loopback device back into the call app. The architecture cleanly separates concerns into an IPC layer (CheetahIPC), a helper process for the browser extension, and the Swift app itself, rather than cramming everything into one process.
The browser extension only works in Firefox and has to be loaded as a temporary add-on each session — there's no Chrome support or packaged install, which is a real gap given most interview platforms (and most candidates) default to Chrome. The build requires manually cloning whisper.cpp as a sibling directory and installing SDL2 by hand; there's no vendoring or submodule pinning, so it'll silently break as whisper.cpp's API moves. It's hardcoded to OpenAI/GPT-4 with no provider abstraction and no cost or rate-limit handling around a key you supply yourself. No commits since January 2025, so a project all about keeping up with fast-moving model and platform changes is effectively frozen. And beyond the maturity issues, using this in an actual interview rather than practice is the kind of thing that gets candidates blacklisted if caught — that's a real risk for anyone tempted to use it as intended rather than for mock practice.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →