Writing
Notes I've actually finished. Pipelines and notebooks when they're ready to read.
Writeups
When you squeeze a webpage, do the boxes move in a pattern?
Sep 2026
I wanted to know if a webpage has any physics. When you make the window narrower, do the boxes move in a pattern a model can learn, or is every page a mess? I measured 3,056 boxes on 25 pages. A one-line CSS rule beat a simple linear model. A tree model beat the rule. Median error went from 26.7 pixels to 7.3.
A Pokémon card agent that plays itself
Sep 2026
I built an agent that plays Pokémon TCG in the cabt simulator. I mined daily match logs for winning decks, then made the agent play itself. Best ladder score 708. Rank 2057 of 6807.
Guess the rock along a drill path
Aug 2026
ROGII wellbore geology. I guessed TVT, the depth marker along a horizontal well, with LightGBM, CatBoost, and Ridge on 5 folds grouped by well. Best of my accepted rows 13.894. Rank 4696 of 6125.
Teach the steps, not the last number
June 2026
NVIDIA Nemotron reasoning. Each puzzle hides a tiny program. I fitted that program, wrote gold step-by-step traces from it, then LoRA-tuned Nemotron-3-Nano. Public 0.704, private 0.720. Rank 2588 of 4182.
A math model that runs on the same machine
May 2026
AIMO Progress Prize 3. I ran gpt-oss on a local vLLM server with Harmony prompts, and I had a Qwen2.5-Math pipeline next to it. Private 38.5. Rank 2345 of 4138.
Will they pit on the next lap?
May 2026
Playground Series S6E5. I mixed CatBoost, XGBoost, and LightGBM over 10 folds to guess whether a driver pits next lap. Private AUC 0.950. Rank 1090 of 3022.
Notes
The floor is worth raising
Aug 2026
I read Taste Labs' research request the week it went up. This is me answering it in public so I can't pretend I didn't.
Taste is the thing we actually hire
Aug 2026
A chair that holds you and a chair you want in the room are different jobs. We keep scoring the first one.
If it only knows the data, put taste in the data
Aug 2026
A model is what you feed it. We keep feeding it the average of the web and then asking why it has no pulse.