Independent AI research, run on my own hardware. Small generative models taken apart and put back together — each experiment ships an illustrated explainer, and most come with a hands-on demo you can poke at in the browser.
An illustrated atlas of a generative model that writes geography as a program — a 33.7M-parameter transformer expanding a map one 500 m tile at a time, with continuity across every shared edge enforced by contract rather than hoped for. Grow a world yourself inside the page.
Flood segmentation from Sentinel-1 SAR with the TerraMind foundation model, fine-tuned on Sen1Floods11. A staged weak-label pipeline, real IoU numbers — and an honest look at why 4,160 free labels made things worse out of distribution.
A DDPM trained on CIFAR-10 in ~250 lines of plain PyTorch — 18.7M-parameter U-Net, linear noise schedule, DDIM sampler, EMA weights — plus an ONNX export verified against PyTorch for in-browser WebGPU inference.
Training a language model end to end on a single GPU — tokenizer, architecture, multi-hour resumable training runs, and what the loss curve actually teaches you.
A source-grounded mission research assistant, with inspectable PDF citations and an honest record of retrieval and answer failures.
A fine-tuned YOLO detector that picks out people and their safety equipment on construction sites — helmets, hi-vis vests, gloves — and flags when someone is working without them.
// 7 experiments logged — more are compiling. build notes live in /notes ▸