Truefit.ai

Voice interviews need audio that survives real network jitter — a naive WebSocket audio pipe means every dropped or late packet blocks everything behind it in the same stream (head-of-line blocking), which is exactly the failure mode you can't have mid-interview.

Truefit splits the session in two: a WebSocket channel carries session signaling — state transitions, tool calls, control messages — while all audio moves over WebRTC/SRTP. SRTP is UDP-based and designed to tolerate loss and reorder without stalling the stream, so a bad packet degrades quality for an instant instead of freezing the whole call. That split was a deliberate trade — more moving parts to run and debug, in exchange for audio that stays responsive under real conditions instead of degrading gracefully in a lab demo and badly in production.

Before launch, interview sessions would occasionally lock into an infinite response loop — the agent kept responding to itself. Two independent bugs were compounding:

• A greedy queue drain. The audio consumer was pulling everything available off the queue in one pass, collapsing up to 40 discrete audio chunks into a single call to the model.
• A clock reset. Frame pacing was derived from a timestamp that got reset mid-session, so the pacing logic computed nonsensical intervals for those merged chunks.

Together, the merged chunk landed at the model with corrupted timing, produced a response that re-entered the same consumer path, and the loop fed itself. Neither bug alone reproduced it reliably, which is what made it expensive to isolate — it only showed up when queue backlog and a clock boundary lined up at the same time. The fix was to bound the drain to a fixed chunk budget per cycle and re-derive pacing from a monotonic clock instead of one that could be reset mid-flight — closing both failure paths rather than papering over the symptom.

The interview agent is structured as a tool-calling, domain-driven state machine sitting behind abstract ports — the agent logic doesn't know or care whether it's talking to Gemini Live or OpenAI's realtime API underneath. Swapping providers is a port implementation, not a rewrite of the interview logic, scoring, or conversation state.

A custom AudioBridge handles PCM resampling and asyncio-based concurrency, so a single session can manage multiple audio participants without the state machine having to reason about the transport layer at all.