Event
LA Hacks
Team
Marc Da Silva, Advaith Karthikeyan, Jordan Joelson
IMPULSE is a real-time incident response platform built at LA Hacks 2026 that turns any iPhone into a smart body camera and gives dispatchers a live, intelligent picture of what is happening on the ground. Built by Marc Da Silva, Advaith Karthikeyan and Jordan Joelson, IMPULSE won the Figma Make Challenge by replacing radio and clipboard with unified AI-powered command intelligence for emergency responders.
What it solves
When an emergency call comes in, seconds matter. But the tools dispatchers and field commanders use today are shockingly fragmented: separate screens for radio, maps and camera feeds with no way to search what happened five minutes ago. Cloud AI adds 800 to 2,400ms of round-trip latency, far too slow for life-safety decisions. Traditional bodycam systems have also been nearly impossible to deploy due to HIPAA and consent law complexity. Edge AI sidesteps all of this by processing footage entirely on-device so sensitive patient data never leaves the camera.
What it does
For field responders, IMPULSE turns the iPhone into a wearable smart bodycam running YOLO object detection and audio analysis entirely on-device. For dispatchers, a live dashboard shows every responder's camera feed with AI outputs alongside it, a prioritized incident alert queue overlaid on a 3D map, and semantic video search so they can type "person on the ground" and get timestamped clips back instantly. A 3D scene reconstruction tool builds an interactive point cloud of any space from field footage.
How it works
IMPULSE uses ZETIC Melange to run YOLO object detection and Qwen2.5 audio reasoning entirely on Apple chips with no server round-trips. A two-stage memory-swap architecture loads the audio encoder and LLM decoder separately to fit within phone RAM constraints. Video frames stream to an on-site ASUS GX10 hub over WebRTC. Semantic video search uses GPT-4o-mini captions embedded into PostgreSQL with pgvector for HNSW cosine similarity search. COLMAP handles 3D scene reconstruction from field footage. No cloud dependency required for the core response loop.