
Kwik 112 is an AI call-taker for India's 112 emergency line (India's equivalent of 911), built entirely on the AssemblyAI Voice Agent API. When a caller panics and switches between Hindi, Hinglish and English, it stays calm, asks one question at a time, and hands a human dispatcher a ready-to-act, severity-graded incident card. ASSEMBLYAI INTEGRATION: One WebSocket runs the whole call Universal-3 speech-to-text, LLM agent replies, turn-taking with barge-in, and 43-term keyterm biasing for Hindi medical vocabulary and Delhi localities. A mid-call JSON-Schema tool, propose_incident_update, structures the agent's read of the emergency; optional severity refinement runs through the AssemblyAI LLM Gateway. ORIGINALITY: the escalate-only floor: A deterministic rules engine sets a severity floor in code. The agent can raise severity above it, but can never lower it, a no-downgrade guarantee enforced by tests, not a prompt. Demo: the agent proposes LOW, the console holds CRITICAL, and a banner shows the override. It is prompt-injection resistant, with a committed test. EVIDENCE: On a reproducible synthetic benchmark it misses 0 of 9 critical calls (100% critical recall); over-escalation is bounded at 16.7%, so it is not a blanket-CRITICAL alarm. The local grade lands in milliseconds (p95 ~5ms), before any model replies. 284 tests pass in CI. HUMAN-IN-COMMAND: A human dispatcher makes every dispatch decision across three checkpoints INTAKE, DISPATCH, RESOLUTION with a full audit trail. The AI assists; it never dispatches. TRY IT: Live demo at kwik112-assemblyai.vercel.app, public GitHub repo, one-command benchmark. Synthetic data only; not connected to any live 112 line.
30 Sep 2026