Things I have built

Zoroaster

Researcher & developer

A model that reads English Wikipedia's edit history and tries to call what the world does next, as a dated prophecy published every night.

The first stage ranks pages. Trained on 133 million revisions across 9.7 million pages — three years of edit history, some 22GB of dumps streamed a page at a time so memory never depends on the size of the corpus — a gradient-boosted model scores every recently edited page by how likely it is to burst tomorrow: far more edits than usual, from at least two separate editors. On the held-out months 74 of its top 100 pages drew a lasting edit the next day, against 58 for the best simple rule, and in a live backtest about one in five of each day's top 100 burst as predicted, twice what ranking by yesterday's edit count manages. Every feature is strictly point-in-time, computed from days before the one being scored — a rule worth a test of its own, since an early version leaked the future through its burst scores and flattered itself nicely.

The second stage writes. Qwen2.5-0.5B and 1.5B, fine-tuned with QLoRA on a single 8GB card, learn to produce the text an edit actually inserted, given the page, the date, and which of its linked pages were bursting the day before. Fine-tuning cuts the perplexity of real inserted text from 15.6 to 6.6 at the smaller size and 5.3 at the larger. The useful negative result: what helps is the neighbouring page's name — a singer's page gaining "She performed at the 2026 Winter Olympics opening ceremony" — while summarising what actually changed on that neighbour adds nothing measurable.

The site is the part that can be wrong in public. Each night a 7B model reads the day's top pages and the week's stories from Wikipedia's current events portal as they stood at midnight, resolves every date relative to the day ahead, and writes at most ten dated predictions. Guardrails drop any that name an individual, that have an organization doing or suffering harm, or that merely restate their evidence; one that invents a number its sources never gave goes back to be rewritten. Everything is labelled a machine-generated forecast rather than news, a judge grades each prediction on its due date (not yet published, while it is checked against hand grades), and a "how the prophet works" page shows the ranked pages beside how the previous day's list actually turned out.

• Python
• LightGBM
• PyTorch
• QLoRA
• Parquet

Read today's prophecy Source

Stele

Designer & developer

A handout generator for tabletop games. Type your text, pick a material, and get back a weathered inscription or an aged document — chipped marble, scorched paper, faded runes.

Start from a medium — marble, sandstone, granite, slate, clay tablet, bronze plaque, wood, paper, parchment, papyrus — then choose a colour, a shape, and a hand to write in, anywhere from Roman capitals and blackletter to copperplate, typewriter, runes, and cuneiform.

The damage is the point. One dial ages the whole piece, or you can layer specific kinds: chips, cracks, lichen, verdigris, burns, water stains, ink blots, tears, folds. Fade eats into the writing itself, and a brush lets you paint wear exactly where you want it. Wrap words in [[…]] to mark them for destruction or {{…}} to keep damage off them, so the unreadable parts of a handout are the parts you chose.

It all renders in the browser: the weathering is GLSL shader work over a WebGL canvas, with no image library and no server behind it. Finished pieces export as 300 DPI PNGs — transparent, if they are headed for a virtual tabletop — or as multi-page ZIPs, full-size PDFs, or a shareable link.

• TypeScript
• WebGL
• GLSL
• Vite

Try it Source DM Tools shelf

Sator

Designer & developer

Anagrams for tabletop puzzles: scramble a password, a name, or a clue, hand the players the letters, and make them work it back.

Scrambles come at a difficulty. Easy keeps the words and their first letters; medium moves every letter; hard runs the words together and parts old neighbours, so no two letters that sat side by side still do. They can be reshaped into a set number of words or a pattern of lengths like 3-4-3, and made pronounceable enough to read aloud — RATIONSESTARDARDISED… rather than NKLEEOESFIITTSARHOE… When the letters cannot do what you asked, because AAB has nowhere to put its second A, it says so and shows the closest it can manage.

It will also find phrases of real words that use every letter — DORMITORY gives DIRTY ROOM — out of a word list you pick plus any names from your own game, or help you write an anagram by hand, showing the letters still unused and the words that would fit in them.

Any line then becomes a puzzle. The card measures how much the clue gives away — letters left in place, old neighbours, pieces of the answer still whole — and lists the other answers the same letters spell, so you can decide which of them you would accept. Add a riddle, decide how many hints the players may take, then show them the clue in large tiles, send a player link that checks guesses without holding the answer, print tiles to cut up and scatter, or pass it to Stele to be carved in granite. The whole puzzle lives in the URL, compressed, so the page never spells the password out during a screen share and a puzzle built before the session comes back with the link.

• TypeScript
• React
• Vite

Try it Source DM Tools shelf

Interactive Graphics Demo

Programmer

A hands-on WebGL scene for poking at computer graphics fundamentals — every part of it is adjustable from the page.

The scene supports a range of primitives that can be added, scaled, rotated, and translated freely. Objects can be parented to one another, so changing a rotation speed moves an entire hierarchy at once.

It also implements the Blinn-Phong lighting model with two light sources, while the stars behind the scene read as a third. Objects carry textures, bump maps, and in some cases animated textures, all swappable at runtime.

• WebGL
• JavaScript

Try the demo Source

Artificial Frame Generator

Software developer

Software that invents the frames between frames, interpolating new video from the footage around it.

The generator estimates interstitial frames from neighboring ones using Farneback optical flow, Canny edge detection, and motion blur.

Testing it was pleasantly self-checking: pull frames out of an existing clip, regenerate them, then compare the reconstruction against the originals mathematically.

• OpenCV
• Optical flow

Source

RPG MegaMart v2

Software developer

A live web app where tabletop players design magic items, publish them, and sell them to each other in shared stores.

Anyone can build items for their game — Dungeons & Dragons, Monster of the Week, Uncharted Worlds — customize them fully, and make them public for other groups to browse. Gather the items you want into a store, add your players, and let them shop.

The first version of this was a PHP prototype running off a laptop on my local network. V2 rebuilt it properly on .NET Core with an MVC structure and a PostgreSQL database.

• .NET Core
• PostgreSQL
• MVC

Visit the site

The Juice Press

Software developer

A satire publication with a stable of writers and, yes, a custom quiz builder.

The Juice Press gives contributors a place to draft, edit, and publish satirical articles and quizzes, with the quiz format built from scratch rather than bolted on.

Visit the site

Robot Vision with OpenCV

Programmer & team lead

A run of projects on teaching a robot to see: image manipulation from scratch, then object detection, then robots acting on what they find.

The groundwork was writing the algorithms by hand in Java — rotations, translations, blurs, masks, edge detection, along with color correction, detection, equalization, grayscaling, and histograms.

From there we moved to OpenCV in Python. First color detection with runtime-adjustable HSV tolerances, then differencing a rolling average of recent frames to pick out motion as it happened. Then object detection with Canny edges and contours, which we tested by counting Skittles and M&Ms by color — genuinely hard when a candy matches the background or overlaps another of the same color, and a useful lesson in how much tolerance tuning real vision takes.

The last stretch put it on hardware in MSU's robotics lab. Blur, edges, and thresholding got a robot following a solid line quickly; a dotted line took adding center-of-mass tracking of the filtered objects in frame, and a 180° sweep at the end of the trail to confirm it had actually finished. Next it searched for a human face, adjusted its distance until it was comfortable conversational range, greeted the person, and tracked their face — restarting the hunt if they left frame for more than fifteen seconds. The final program combined everything: navigate an obstacle course by distinguishing obstacles from safe path, find a person, take an item from them, and retrace the route.

• Python
• OpenCV
• Java

Image processing I Image processing II Line following Face finding