Capture
React streams PCM audio through an authenticated WebSocket.
Voice AI · Adaptive systems / 2026
InterviewLab
Building an adaptive interviewer without giving the model unrestricted control.
01 / The problem
Complexity should live in the system,not in the experience.
A useful mock interviewer must react to the candidate, but an autonomous model can repeat itself, skip rubric coverage, or close unpredictably.
02 / System flow
From input to a reliable outcome.
Transcribe
Deepgram provides speech-to-text, the conversational loop, and text-to-speech.
Evaluate
A LangGraph director evaluates complete answer turns and owns interview state.
Report
PostgreSQL stores transcripts, assessments, coverage, and the final report.
03 / Decisions that shaped the build
01Engineering choice
LLM judgment, deterministic control
The model scores answer evidence; explicit code decides follow-up, harder, new-topic, or wrap-up routing.
02Engineering choice
End-of-turn over fixed timers
Deepgram thinking events flush pause-separated transcript fragments as one answer before evaluation.
03Engineering choice
Single voice WebSocket
Native barge-in and one round trip were prioritized over independently orchestrating STT, LLM, and TTS.
04 / Reliability
Designed for imperfect conditions.
- Autonomous voice-agent fallback when the graph cannot start
- Strict report schema with normalization
- Assessment-derived report fallback
- Soft wrap, graceful close, and hard duration ceiling
05 / Reflection
“The strongest architecture boundary was letting AI judge meaning while ordinary code controlled sequence, limits, and recoverability.”Explore selected work