Limeade

The safety layer Lime is missing.

Comment

Inspiration

Electric scooter riders in cities like Chicago face serious safety hazards like potholes, aggressive drivers, construction zones with zero advance warning. Riders have no visibility of threats approaching from behind or peripheral directions, and current scooter platforms like Lime offer no built-in safety intelligence. In 2023, Chicago saw over 1,000 e-scooter related injuries. What inspired us to pursue this project was one of our teammates' sister, who fell off a lime scooter and was severely injured.

We asked "What if the scooter itself could see the road and protect the rider?" Limeade is designed as a native hardware safety module built directly into Lime scooters, using a forward-facing camera, onboard processing, and haptic feedback through the handlebars, all invisible to the rider.

What it does

Limeade is an embedded AI safety system for Lime scooters. A camera module built into the scooter's handlebar assembly continuously watches the road ahead. A multi-stage AI pipeline processes each frame in real-time:

  1. YOLOv8 computer vision detects vehicles, potholes, and road infrastructure
  2. Object tracking + optical flow estimates approach rates and whether objects are on a collision course
  3. Claude Haiku vision receives the frame alongside structured detection data and reasons about whether it's a genuine threat, filtering out parked cars, parallel traffic, and false positives
  4. Claude decides the exact response: the rider feels a haptic pulse through the handlebars calibrated to urgency, and hears a spoken warning like "Pothole ahead on your right, steer left" through the scooter's built-in speaker
  5. Confirmed hazards are saved to a crowdsourced database with GPS and the camera frame, so the next scooter approaching that same location gets a proximity warning before the camera even sees it
  6. Fleet operators monitor all scooters via a live dashboard, an interactive map with camera thumbnails at each hazard, AI threat assessments, and a real-time event feed

For our hackathon prototype, a phone mounted on the handlebars simulates the embedded hardware module, capturing frames, delivering haptic/audio alerts, and streaming GPS. The backend pipeline, AI reasoning, and fleet dashboard are fully functional.

How we built it

Phone: Expo React Native app. Streams camera frames over WebSocket to a laptop on the same WiFi. One frame at a time, waits for ack before sending the next. Server: FastAPI WebSocket server on the laptop. Every frame runs through two YOLO models (vehicles and potholes), IoU tracking across frames, and Farneback optical flow to measure real motion. This gives us a danger score per frame. Trigger logic: instead of calling Claude on every frame, we gate it: danger must stay above 0.5 for 3 consecutive frames with a 3-second cooldown between calls. This keeps latency low and API costs minimal. Claude Haiku: when the gate passes, we send the camera image plus all detection data and a 3-frame trend to Claude. It reasons whether this is a real threat or a false positive and returns a structured response with urgency level, a one-sentence warning, and a haptic pattern. Alert: phone receives Claude's response, vibrates in a pattern based on urgency, and speaks the threat summary out loud via text-to-speech. Database: every pothole detection and hazard is logged to PostgreSQL with GPS coordinates and a timestamp.

Challenges we ran into

False positives were our biggest technical challenge. Distant objects and not dangerous all triggered alerts early on, making the system unusable. We had to build a multi-stage filtering pipeline with CV detection, then rule-based danger scoring, then Claude reasoning just to get alerts that actually meant something. Even then, tuning the balance between catching real threats and ignoring noise something we consistently worked on

Accomplishments that we're proud of

We are proud that we are able to make the roads a safer place to navigate around with an electric scooter.

What we learned

How to use different tech stacks and tools we weren't familiar with, such as computer vision models, front-end libraries, web sockets, and app development with react native.

What's next for Limeade

We want to partner with Lime to move from our phone-based prototype to dedicated hardware which we envision to be an embedded camera module and haptic motor built into the handlebar assembly, invisible to the rider. We'd also train a Chicago-specific hazard model on local road conditions (construction plates, raised manhole covers, uneven L-track crossings) and add route planning that steers riders away from hazard clusters. Long-term, aggregated hazard data could be shared with the city's infrastructure teams to prioritize road repairs where riders need them most.

Built With

Share this project:

Updates