Projects

Model Arena

Ten small tasks; three or four models race head-to-head. Scoring is deterministic — correct → cheapest → fastest — so the winner is arithmetic, not vibes. Replays are recorded and free; live runs stream real tokens behind a bot-check.

One stream feeds every lane at once; scoring is a fixed rule; live spend stops at $0.50 a day.

Next: visitor-submitted challenges, moderated, with the same scoring discipline.

Guess Which AI

A daily parlor game: one odd prompt, three anonymous answers — on some days a fourth written by me. Spot who wrote what. The reveal prints each answer’s model, capture date, and exact price.

Answers recorded once, first take kept; sets chosen by rules written down in advance; nothing runs server-side.

Next: a season-two roster swap once the current models’ tells get famous.

Token Receipt

The tokenizer under this site’s hero. Exact OpenAI counts computed in your browser via gpt-tokenizer (o200k); Claude counts shown as labeled estimates because Claude’s tokenizer is private. Nothing you type leaves the page.

Client-side compute, a lazy-loaded tokenizer bundle, estimates labeled as estimates.

Next: side-by-side cost diffing for two prompts.

Linework

Daily puzzle games for writing craft — the shape and ritual of the NYT word games, but the skill being exercised is prose. Started on a bet that as AI drafts more of everyone’s words, deliberate practice for human writing gets more valuable, not less.

Next: finish the first three apps and grow the excerpt bank to launch depth.

Pricing Optimizer

A simulate-and-compare tool for pricing decisions. The chart wiring came out genuinely sleek. The demand data underneath was generated, so the answers couldn’t mean much — and that’s why it stopped.

Next: a rebuilt “Pricing Explorer” — same domain, aimed at exploring and comparing pricing models rather than optimizing one.

SEC Narrative Drift

Diffs the story a company tells between SEC filings — which risk paragraphs appeared, vanished, or quietly changed adjectives.

Next: parked. The extraction practice was the real yield; the year-over-year diffs were mostly noise, and that lesson cost three months.

Response Option Fit

Checks whether a survey’s answer options actually fit its question — how survey data goes wrong before anyone answers.

Next: none planned — it did its job as a study.