Tally: Reliability Layer for Voice Agents

Voice agents trust their own tool calls. A customer corrects an order mid-sentence, "two burgers, no wait, make it three," and the agent either mishears it, talks over it, or confidently commits the wrong quantity while sounding completely sure of itself. Nobody finds out until the order, the booking, or the total is wrong. Tally is a reliability layer that sits between a voice agent and the action it's about to take, built on AssemblyAI's Voice Agent API. The loop: an independent speech stream runs alongside the agent's own recognizer, giving Tally its own evidence rather than trusting the agent's transcript. Every tool call is validated against that evidence before it commits. If the two disagree, or evidence hasn't arrived yet, the call is held, not silently allowed. Instead of failing the whole order, Tally triggers one scoped, targeted re-ask about just the disputed item. Once confirmed, the order commits correctly. Every caught failure becomes a stored, replayable case, and a config change can't go live until it passes the full regression suite built from real failures. This isn't a mock. The full loop, correction detection, hold, targeted repair, barge-in handling, and spoken-drift correction, has been validated live against AssemblyAI's real managed Voice Agent and real human speech. A full real-speech validation pass across 3 speakers and 36 recordings found and fixed two real bugs in the correction-parsing logic, and honestly documented the limitations that remained rather than hiding them. Try the live guest demo with no login required, place a demo order, and watch the evidence timeline flip from CONFLICT to ALLOWED in real time.