I ship AI systems and then try to break them: a language model running bare metal on an FPGA, eval harnesses that run the code instead of asking a model, and agent fleets whose planner prices the job before anything spawns.
LinkedIn · Every original repo
| Shipped | Project | What it is |
|---|---|---|
| Nov 15 | hbs-eta-2026 | Notes from the HBS ETA Conference 2026: Romney, the new SBA rules, Guild Garage and roll-ups, in plain language. |
| Nov 9 | thinkstop | Stop a reasoning model's thinking loop and ask for the answer. |
| Nov 8 | sundai-hack-148 · live | DEFY: an AI seals its guess before each of your 12 taps, then dares you to pick anything else. Sundai Hack 148. |
| Nov 7 | leveraged-buyouts-explained · live | Leveraged buyouts, explained: search funds and ETA in plain language. |
| Nov 7 | memory-dream | Read-only audit of Claude Code's memory folders: reports what a clean-up pass would fix and prints a proposed plan. |
My five newest portfolio repos, pulled from the GitHub API every morning.
| Project | What it is | |
|---|---|---|
| 🔌 | vibe-silicon | We configured Intel's prebuilt Nios II soft core into an Intel MAX 10 FPGA (DE10-Lite) that had no processor, then ran llama2.c on it bare metal. |
| 🧪 | smoleval | Eval harness for small code models: HumanEval + MBPP, sandboxed execution, unbiased pass@k, held-out set locking. |
| 🧠 | reasoning-over-recall | Why you'd train a small language model to reason instead of memorize - a plain-language explainer checked against the phi-1 paper, plus a working… |
| 🕸️ | graph-engineering-kit | The runnable half of graph engineering. A planner that prices a job before spawning it, ten rules that stop the three failure modes, four production… |
| ⚡ | local-llm-decode-ceiling | Why local LLM decoding on Apple Silicon reaches only half its memory-bandwidth ceiling, and what speculative decoding wins back. |
| 💧 | why-ai-needs-water · live | A 3D liquid-cooled GPU server loop you can take apart in your browser. One input drives every gauge. |
How this profile builds itself
- The portrait is my headshot turned into type by
scripts/ascii_portrait.py: background removed, local contrast lifted, then every cell mapped onto a 20-step ramp of characters, mostly letters, sorted by how much ink each one puts down. The photo never enters the repo, only the letters. - Easter egg: the mid-tones of that ramp are
l i s u n o w, the letters of my name, so wherever the photo sits in the mid-greys, my name does the drawing. In the light card that is most of my face. - The card is plain SVG written by
scripts/build_card.py, Python standard library only. Every number on it comes from the GitHub GraphQL API. Thefortuneline quotes one of my own repo descriptions, a different one each day. - The schedule is one GitHub Action that runs daily on the built-in token, so there is no personal token to expire, and it commits as the Actions bot, so it never pads my contribution graph.
- Credits: the snake is Platane/snk. The portrait-plus-neofetch idea comes from Andrew6rant; this is a from-scratch take on it.