Diyar Polat

A constraint solver for the Kanoodle puzzle: twelve irregular pieces, each with eight orientations, that must tile a board exactly. Written in Haskell as a backtracking search with pruning, wrapped in a small HTTP API and a browser UI you can play with directly.

The half I find more interesting is how it's operated. It runs on an NVIDIA Jetson Xavier NX on my desk, under a systemd unit that gives it nothing it doesn't need: DynamicUser so it runs as a transient unprivileged account with no home and no login, ProtectSystem=strict and ReadOnlyPaths for a read-only filesystem, NoNewPrivileges, plus CPUQuota and MemoryMax caps so one expensive search can't starve the board. A timer polls the release branch every five minutes and promotes a new build only if it compiles: a failed build leaves the running version untouched rather than taking the site down. The board exposes no open ports at all: a Cloudflare tunnel dials outward, so nothing on my home network is reachable from the internet.

Counting Focus Forming Units in fluorescence microscopy images is slow and inconsistent when done by hand. This OpenCV pipeline does it automatically: preprocessing and thresholding to separate foci from uneven background illumination, then feature extraction and counting, with a reproducible evaluation workflow so results can be checked against manual counts rather than trusted blindly.