AI-Assisted Puzzle Solver
01A computer-vision app that scans a physical puzzle, rebuilds it digitally, and gives progressive solving hints — running entirely in the browser.
AI-Assisted Puzzle Solver
My internship project at Smart Toys & Games (Smart NV): a computer-vision app that scans a physical puzzle through the phone camera, detects every piece, rebuilds the board digitally, and provides progressive hints — from a single cell up to the full solution. Everything runs on the user's device, with no backend and no installation.
My contribution
I built the whole pipeline: a dataset of 1,800+ synthetic images rendered in Blender plus 240+ real photos hand-labeled in Roboflow; a YOLOv26 Nano model trained in two stages (synthetic, then finetuned on real) reaching 99.5% mAP; perspective correction with homography and grid mapping; and a backtracking solver with progressive hints. I then ported the entire CV pipeline from Python/OpenCV to pure JavaScript, running the model in the browser via ONNX Runtime (WebAssembly), and reused the same architecture for a second puzzle.
What I learned
This was my first AI project in a professional environment, and the result will actually be used by children playing Smart NV puzzles. I grew a lot in computer vision, model training, data generation, and building full applications from scratch — and, working alongside senior developers, I learned when to push through a challenge and when to speak up and redirect.