911 DispatchLab

Our agent simulates live 911 emergency calls in the browser, with an AI “victim” voice powered by ElevenLabs and a human trainee acting as the operator. It handles real-time two-way audio, live transcription, and a protocol coach that nudges the operator through the first six critical questions (location, emergency type, injuries, number of people, immediate danger, identity/callback).

The working prototype is already stable for single-operator sessions: calls can be started, audio streamed, transcripts rendered, and protocol steps tracked without manual intervention. Technical complexity sits in coordinating multimodal streams and tools: ElevenLabs TTS/STT, an LLM for scoring, Web Audio for mic and playback, and real-time state sync in the UI. Innovation comes from treating the “call” itself as an autonomous training agent that adapts caller behavior, gives partial or confused answers, and produces structured feedback after each session, not just raw audio. In real-world use, this could reduce training cost, standardize early-call behavior, and measurably improve response quality for emergency centers.

Theme alignment: the project literally turns browsers, voices, clouds, and tools into one cohesive agent. The browser hosts the operator console and orchestrates user input (mic + text). Voices are generated and transcribed through ElevenLabs in the cloud. An LLM coordinates the scenario logic, caller behavior, and feedback. The tool layer includes protocol tracking, scoring, and dashboards that act on top of transcripts and call metadata. Together they behave as a single training agent that listens, speaks, reasons, and evaluates.

Tech stack and tools used:

Partnered Technologies: Bolt.new and ElevenLabs
Frontend: TypeScript, Vite
UI & styling: Tailwind CSS, custom CSS animations
Audio: Web Audio API for mic capture and playback
AI voice & audio intelligence:
ElevenLabs Text-to-Speech
ElevenLabs Speech-to-Text
ElevenLabs voice IDs for personas
LLM (for logic, prompting): Gemini (configurable)

Storage & persistence: Supabase for frontend and SQLite for backend
Agentic Framework: LangGraph (orchestrates both Gemini and ElevenLabs SDK)