~/projects/strativ-greenroom
Strativ AB · internal product, in active use for candidate screening
AI interviewer in active use for candidate screening. Candidates answer spoken questions, the system transcribes them live, adapts its follow-ups and produces an assessment for recruiters.
Greenroom runs a structured, spoken screening interview. Each answer is transcribed in real time, summarised by an LLM and fed into a phase machine that decides between a follow-up and the next question. Recordings upload in the background throughout the session, and after the interview a job pipeline produces a full transcript and an assessment for the recruiter.
Candidate audio streams as 16 kHz mono PCM directly from the browser to AssemblyAI. Django only mints a short-lived token per question, so the vendor key never leaves the server and audio never passes through the app servers.
After every answer an LLM summarises it and a phase machine (warm-up, technical, wrap-up) decides whether to ask one follow-up or move on. Question quotas per phase come from defaults or from questions prepared for that role.
Answers are accepted immediately (202) and processed off the request path. The next question is pushed to the candidate over WebSocket through the Redis channel layer.
MediaRecorder uploads 30-second chunks in parallel with the interview. Failed uploads are buffered in IndexedDB and retried every 15 seconds, with a grace window after the session ends.
Background jobs finalise the recording, produce a full-session Whisper transcript and generate the assessment, with retries and a manual-review fallback when generation keeps failing.
Two application processes share PostgreSQL and Redis. Audio goes from the browser to the speech-to-text provider directly.
One cycle per question, repeated until the phase machine finishes the interview.
Each phase has a question quota and allows at most one follow-up per question.