KODA Vision

We entered Google's Gemini Hackathon as Koda, a team of five. We started from a concrete question: what does a blind person need to know when walking across a campus they don't know? Where the stairs are, what the sign on the door says, whether a car is coming.

That became KODA Vision: a camera mounted on glasses that describes the surroundings out loud, reads the text it finds, and answers questions like "what's in front of me?" while the user holds a button.

Narrating every second wears people out fast. So the system compares each frame with the previous one and only describes again when the scene actually changed.

Warnings have a hierarchy. A STOP or DANGER sign read by OCR, or a nearby vehicle, cuts off whatever is being said. Obstacles come next, and the general scene description goes last.

YOLOv8 Nano and Tesseract run on the device, so hazard alerts don't depend on the network. Gemini 2.5 Flash gets the frame plus the text OCR already extracted and returns one or two sentences. Google Cloud TTS turns them into Spanish speech, played through bone conduction so the user can still hear the street.

The system also accepts a recorded video instead of the camera. That saved us when testing and debugging without walking around every time.

On April 29, 2026 we presented the prototype live on the Universidad de La Sabana campus in Chía: a laptop, a webcam, and bone-conduction headphones. The system described the surroundings, read the signs along the way, and answered spoken questions.