I built a mini agent harness in JavaScript, fine-tuned Qwen3-0.6B on my RTX 3060 to be its brain, and ran the whole thing over WebGPU — no server, no API keys. The model went from 30% to 77.5% on a prior-auth review task. The GPU crashed four times during training and quantization turned out to be the real boss fight.
I fine-tuned a real model on my RTX 3060 to settle an argument: fine-tuning teaches a model how to behave, not what to know. The model learned a perfect persona — and confidently invented every fact it was asked. Here's the experiment, the four broken adapters it took to get there, and the live demo.
A personal account of reverse-engineering a thermal printer, learning BLE from scratch, and working alongside claude code that compressed weeks of debugging into a single weekend.