← Home
Writing
Engineering and research notes from building agent infrastructure on hardware I own. Some of it ships as a tool, some of it as a paper, and one post is about a script nobody has read in 3,500 years.
How Blackreach works, and why I built it differently
ReAct loops, DOM walking, stealth Playwright, and why every other approach kept failing on real sites.
Multi-agent coordination without an SDK
One port number. All your agents share memory. Zero code changes required.
We spent months measuring whether our AI was getting better or worse
Most teams ship agents without knowing if they actually work. I built a tool to fix that.
Linear A, and why I want to take a crack at it
70 years of failed phonetic matching. Why the tools available now are different.
I gave Claude Code a voice, and why nothing else worked
VoiceMode, ElevenLabs, OpenAI TTS, and why I scrapped all of them and built a local TTS with karaoke highlighting.
I built a GitHub dashboard for my terminal
BubbleTea, Go, and why I never want to open a browser just to check my notifications again.
Fine-tuning a protein model to find new antibiotics
ESM-2 and LoRA, 88.3% F1 on a leakage-free benchmark, 1,980 sequences screened. Plus the experiment that failed and why it mattered.
Honest results from a small GRPO lab
What boundary-SFT did for a 3B model, why the same recipe got rejected on 7B, and why I keep the negative results in.