The problem
Generic interview practice asks about algorithms nobody wrote. Grounding the questions in a specific diff makes them answerable only by someone who understands that change - which is the thing an interview is actually testing.
What I built
- Generated repo-aware questions across four difficulty bands, from junior through staff.
- Accepted answers by voice via Whisper transcription or by text.
- Returned structured feedback scored on clarity, depth, correctness and communication.
- Added pull-request-level analysis so a single change can be interrogated instead of a whole repository.
Stack
TypeScriptReactGroq WhisperLLM APIsGitHub API
Skills applied
- LLM application design
- Speech-to-text integration
- Structured output evaluation
- GitHub API integration