KWIK 112Synthetic dispatch simulationStart live demo call
KWIK 112 dispatcher console showing a cardiac arrest incident on the Delhi situational map

Live emergency simulation

KWIK 112

Use the ring time.
Prepare the human response.

Proposed middleware for the 112 queue. Hindi, Hinglish, and English intake; a safety floor the model cannot lower. The citizen would dial from a keypad phone on 2G—no app, URL, or reading. This browser demo emulates that call.

Built on the AssemblyAI Voice Agent API; optional severity refinement runs through the AssemblyAI LLM Gateway. Independent synthetic demonstration, not an official 112 service.

Start live demo call Open dispatch console

Measured, not implied

Safety starts before dispatch.

Critical recall
100%

9/9 held-out cases · Wilson 95% LB 0.70

Threat detection
100%

3 of 3 held-out cases

Local triage latency
0.042 ms

Median on the benchmark

Development regression suite, 30 synthetic calls: 100% critical recall (9/9) — no critical call missed — with location text 100% (25/25). The full case-level breakdown, including type/severity accuracy and under- and over-triage, is published unedited. Local p95 5.219ms; timings exclude providers and vary by run. Inspect the committed results · Raw evidence.

  • 0.28% genuine of about 16 lakh daily combined 112/Dial 100 calls in Telangana. 25 June 2026 report.
  • Under 15 seconds is the MHA answer-speed target; the reported national response average is about 18 minutes. Different stages, not measured ring time. MHA, 2015, p20; DGP report, 6 August 2026.
  • 10,000 of 15,672 daily Delhi calls were reported blank; IVR filtering was already in place. 1 October 2019 report.
  • A merged queue: SaveLIFE Foundation v. Union of India, W.P.(C) 726/2024, 2026 INSC 567, ordered six helplines integrated into 112 within three months. The window has closed by calendar inference; implementation is not established here. Order, 26 May 2026.

From call to closure

One operational thread. Six decisions.

Every stage adds context without taking control away from the dispatcher.

  1. 01

    Receive

    A caller speaks naturally in Hindi, Hinglish, or English.

  2. 02

    Understand

    Transcript, location, injuries, and voice cues reach the desk.

  3. 03

    Grade

    Deterministic rules assign an immediate safety-first priority.

  4. 04

    Refine

    Model review may escalate severity, but never downgrade it.

  5. 05

    Dispatch

    A human selects the response and sends units en route.

  6. 06

    Audit

    Evidence, overrides, units, and decisions remain traceable.

Three scripted calls

Hear the pressure. Watch the incident form.

Each scripted performance carries a distinct language and emergency through the incident pipeline. Its emotion frames are simulated; live calls run on the AssemblyAI Voice Agent API. Fleet and ETAs are synthetic.

Open voice station
01RameshHinglishMotorcycle collisionA panicked bystander near Moolchand Metro reports two injured riders.02Sharma jiHindi / HinglishCardiac arrestA distressed family caller reports no breathing and no pulse in Shalimar Bagh.03JohnEnglishCommercial fireA market caller reports smoke, trapped people, and collapse risk in Chandni Chowk.

Built for scrutiny

Conservative by code. Accountable by design.

Deterministic triage sets a safety floor. A calm report of no pulse remains critical. Emotion stays within its severity band; possible prank cues are annotations, never automatic rejection. Humans record INTAKE, DISPATCH, and RESOLUTION; overrides need notes. The browser-local audit persists until storage is cleared.

  • 30 held-out synthetic calls
  • 0 autonomous dispatch decisions
  • 1 auditable incident timeline

The fastest way to understand it

Place the call. See the response.

Start live demo call View benchmark

Independent simulation. Not an official 112, ERSS, government, or C-DAC service. It does not connect to emergency infrastructure. Safety context