Evaluate Kwik 112 in 120 seconds

Problem

The problem is a busy intake queue, not a missing citizen app. These sources describe different places, periods, and measures; vehicle response time is not telephone ring time.

• Telangana reporting puts genuine emergencies at about 0.28% of roughly 16 lakh daily calls across the combined 112/Dial 100 system. 25 June 2026 report, citing police data.
• The MHA NERS guidelines target an average answer speed below 15 seconds. Separately, a Telangana DGP statement reports a national emergency response average around 18 minutes. These are different stages of response. MHA guidelines, 2015, page 20; DGP report, 6 August 2026.
• Delhi reported roughly 10,000 blank calls out of 15,672 daily calls shortly after the 112 launch. The report also describes IVR filtering already in place. 1 October 2019 report.
• SaveLIFE Foundation v. Union of India, W.P.(C) 726/2024, order dated 26 May 2026 (2026 INSC 567), directs integration of 100, 101, 102, 108, 1033 and 1091 into 112 within three months. As of this review, that stated window has closed; this is a calendar inference, not proof of nationwide implementation. Official order, 26 May 2026.

Working build

No login, no credentials, no setup: place a call above, watch the transcript, detected language, and deterministic grade appear live, then run the three human checkpoints. Every number on /benchmark regenerates from the repository with one command.

Usability

The proposed citizen interface is the dial pad: a keypad phone on a voice-capable 2G network, no app, URL, reading, or data plan. Hindi, Hinglish, and English today. This browser demonstration emulates that call; it is not connected to the telephone network. The dispatcher receives a pre-graded card.

Product thinking

The live agent proposes an incident through a JSON-Schema tool, but the console owns the record: the AI can raise severity above the deterministic floor, never lower it — enforced in code with regression and prompt-injection tests, not a prompt promise. This is a no-downgrade guarantee on the grade, not a claim of perfect triage.

A calm caller reporting no pulse must still grade critical — the local rules set that floor regardless of tone. Any prank signal is an annotation for human review, never proof of a prank, a downgrade, or automatic removal from the queue. The intended deployment sits before C-DAC NG112 call handling; this build demonstrates that boundary without claiming an integration.

End-to-end thinking

Voice intake → instant rules grade → optional escalate-only refinement → three human checkpoints with written-note overrides → unit reservation, road-following routes, and a session-local audit trail. Every value carries provenance: local rules / model / fallback; measured / simulated / absent.

Honesty

On a 30-call synthetic regression suite the deterministic grader holds 100% critical recall (9/9; Wilson 95% lower bound 0.70) — no critical call is missed — with location text 100% (25/25). Type and severity classification on that corpus is a work in progress; the full case-level breakdown, including under- and over-triage, is published unedited on /benchmark. This is a development regression suite, not clinical or field evidence. Reproduce via npm run evaluate:local; CI repeats the critical-recall gate. The model cannot dispatch, and the small corpus cannot establish real-world safety.

Prior art and the accountable grading boundary

VANKI proved translation belongs on the 112 line. We start where that ends: after the words are understood, something still has to decide how dangerous the call is — and must never quietly decide it is less dangerous than it looks.

Bengaluru's VANKI supports multilingual emergency communication; Telangana's TG-ERSS also reports automated handling and filtering. We credit those deployed systems. The cited reports do not publish an executable severity floor or reproducible grading corpus; that is the evidence Kwik makes inspectable, not a claim that nobody else has safety controls. VANKI, 30 April 2026; TG-ERSS, 6 August 2026.

Why keep a model away from dispatch authority?

A Danish randomized trial found no significant improvement in dispatcher cardiac-arrest recognition in its primary comparison. That is why Kwik grades with rules first and requires human dispatch decisions, rather than treating AI assistance as proven clinical benefit. Blomberg et al., 2021 randomized trial. The human-oversight principles of EU AI Act Article 14 inform the design; this demonstration claims no certification or regulatory compliance.

What is real, simulated, or synthetic — in one table

Every capability of this build, labelled. The same provenance labels appear inside the product interface.

The stack

The live call runs end to end on the AssemblyAI Voice Agent API over a single WebSocket: Universal-3 speech-to-text, agent replies, turn-taking and barge-in, keyterm biasing from the triage lexicon, and a mid-call propose_incident_update tool. The deterministic no-downgrade floor and the evaluation harness behind every number on /benchmark are provider-independent code.

Boundaries

• Kwik 112 is an independent demonstration. It is not an official 112, ERSS, Government of India, or C-DAC service, and it contacts no emergency infrastructure.
• Use synthetic scenarios only. A live call sends microphone audio to the AssemblyAI Voice Agent API; optional refinement sends transcript text to the AssemblyAI LLM Gateway, and address lookup may send location text to a geocoder. Do not enter real personal data. The AI never dispatches; human overrides require a written note.
• The benchmark is a 30-case versioned synthetic corpus with a held-out split — a safety-oriented engineering measurement, not clinical evidence.