← All work
Computer Vision
royale.coach
A real-time game coach that captures a live mobile screen, reads game state with OpenCV and OCR, and renders coaching on a web dashboard.
Real-time CV · no LLM in the hot path
The problem
Coaching a fast real-time game requires reading the screen and reacting within the frame budget — too fast for an LLM call in the loop, and too noisy for naive image matching.
The approach
I built a capture-and-vision pipeline: pull the live device screen over ADB, parse game state with OpenCV and OCR, and stream structured state to a web dashboard — keeping the hot path purely computer-vision so it stays real-time.
What I built
- An ADB screen-capture pipeline
- An OpenCV + OCR state-reading layer tuned for speed
- A Next.js dashboard rendering live coaching
The outcome
A working real-time coaching tool that proves heavy CV work can run inside the frame budget — LLMs reserved for analysis after the fact, not the hot path.
Stack
PythonADB captureOpenCVOCRNext.js