UKUCHA

Problem

In underground mining, detecting hazards and fallen people is critical, but the environment has no stable network — any solution has to run on cheap hardware and in real time, without relying on continuous connectivity.

Constraints

• Cheap hardware (ESP32-CAM).
• No stable connectivity inside the mine.
• Real-time processing on limited compute.

Architecture

The ESP32-CAM captures video and streams it over WebSockets to a FastAPI backend. Two models run there in parallel: YOLOv8n for object/hazard detection, and MediaPipe Pose to detect fallen people.

Decisions

• YOLOv8n (nano) instead of a larger variant, given the compute budget available during the hackathon.
• The project’s biomimicry angle was inspired by APOPO’s HeroRATs and the Bristol Robotics Lab — hence the idea of a system that “inspects” tight spaces the way a trained animal would.

My role

I led a 5-person team. I was responsible for backend development: ingesting sensor data (ESP32-CAM streaming) and wiring it up to the frontend.

Result

Top 6 at the FLIT Hackathon 2026 (Arequipa).