.png&w=256&q=75)
🤓 Latest Submissions
Patchline
Most voice agent demos show the happy path. Patchline is built for the path where AssemblyAI mishears something that matters — an order ID, a refund amount — and a naive agent would just act on the wrong value. Patchline sits between AssemblyAI's real-time speech-to-text stream and a set of business tools, checking every extracted entity against real evidence before any tool (lookup_order, request_refund, update_shipping_address) is allowed to fire. When a value is weak or ambiguous, the Action Gate blocks the call and asks one targeted repair question live, in the same conversational turn-taking — a real evidence-driven follow-up, not a script. The caller's own confirmation, not a model's confidence, is what turns an ambiguous value into a verified one. Every recovered failure is captured as a permanent regression: the audio, the wrong value, the confirmed correct value, and how it was confirmed. Operators can replay any regression against a candidate AssemblyAI config (speech model, keyterms_prompt, prompt, turn-detection settings), see whether it actually fixes that failure, and only promote it once it passes the full regression suite — not just an aggregate score. Nothing is silently "learned" by retraining a model; every improvement is a new, replay-tested config with a full audit trail and one-click rollback. Built on AssemblyAI's Realtime Speech-to-Text API (Universal-Streaming). Read-only tools fire deterministically once an entity is verified — no LLM decides whether to call them, so the highest-risk path has no model in the loop. Groq is used narrowly: as a fallback entity extractor when the transcript alone doesn't cleanly parse, and for TTS via Orpheus. Verified against a real, non-mocked AssemblyAI account — including finding and fixing a real keyterms_prompt wire-format bug that had silently broken every "boosted vocabulary" config, plus a full live walk of block → repair → regression → replay → promotion → first-pass success.
30 Sep 2026
