LookOut

Inspiration Traffic incidents kill over 42,000 people in the U.S. every year, and a major factor is delayed detection — it takes 5–10 minutes on average before anyone even knows a crash happened. Secondary crashes pile up in that window. We asked: what if existing traffic cameras could detect incidents in seconds and reroute drivers instantly?
What it does LookOut analyzes live traffic camera feeds using AI, classifies incidents (accidents, hazards, speed sensors) with a severity rating, and alerts drivers in real time with voice warnings and smart rerouting that avoids every active incident on the map.
How we built it An 8-stage async pipeline: Gemini for multimodal image classification, SQLite for structured incident storage, Actian VectorAI DB for vector embeddings and similarity search, Sphinx as a decision/explanation engine, OSRM for intelligent multi-route scoring, and a Next.js/React/Leaflet frontend with real-time alerts and Web Speech API voice warnings. We also trained a YOLOv8 model as a swappable, cheaper perception layer for edge deployment at scale.
Accomplishments that we're proud of The rerouting actually works — it scores multiple route candidates and picks the shortest path that clears every incident. Gemini classifies raw camera frames into structured, actionable data in a single API call with no training data. And every decision in the pipeline is explainable through Sphinx.
What we learned OSRM's public server rate-limits aggressively — we learned to send sequential requests with delays instead of parallel blasts. We also learned that balancing route avoidance with efficiency is critical; maximizing distance from incidents without a duration penalty produces absurd detours.
What's next for LookOut Live camera integration (RTSP streams), two-tier inference (YOLOv8 screens every frame, Gemini classifies flagged ones), push notifications to nearby drivers via mobile, and city-level dashboards for DOT teams to monitor incident hotspots and response times.