AI Interviewer
· Python · FastAPI · LangGraph · PostgreSQL · Redis · WebRTC · LiveKit
Built with a team for AI Product Hack.
The interesting problem was not generating the report — it was that the interview could not end until the report was ready, which made every session hostage to the slowest model call. I moved report generation into background jobs with retries and a reconciler that finds and restarts work orphaned by a crash, so finishing the interview and producing the write-up stopped being the same event.
Around that: a competency profile scored 0–10 with weighted blocks, a draft of the candidate’s feedback that a human confirms rather than writes, OpenAI ASR/TTS adapters behind a config, and an LLM client that caches the structured output call.