Known issues / current state

Voice briefly stutters at the start of every reply. Root cause: Render's free tier gives only 0.1 CPU, and real-time TTS synthesis falls behind briefly before catching up — visible in logs as "inference is slower than realtime". Not a memory issue; a CPU throttling issue.

Out of Memory ( OOM ) recurrence — VAD was re-enabled at some point after the memory fix, which may reintroduce the earlier out-of-memory crash under real conversation load.

First click after idle is slow — direct trade-off of num_idle_processes=0; nothing is pre-warmed, so first connection after a period of inactivity has to cold-start the full pipeline.

Gemini Api rate limiting
Mitigation in place (graceful fallback message) but not a real fix. A same-day fallback option: swap for openai or anthropic