A working portfolio of 33 deployed applications.

A GPT written from scratch in TypeScript with no dependencies — tokeniser, self-attention, hand-derived backpropagation and AdamW all explicit — trained and sampled in a Web Worker, so the tab never blocks and no prompt leaves the browser.

Retrieval that shows its work: the document is chunked, embedded and searched, and every answer cites the passages it used beside the dates, places and people it pulled out. Inference on Cloudflare Workers AI, so the answer can be checked.

A search-based controller rather than a gradient-trained one: each attempt is scored, the search runs in a Web Worker, and the page plots what each generation achieved. The first 360 generations fail on the published score.

An image classifier trained in the browser on labels you supply: a Web Worker trains so the tab stays responsive, weights persist in IndexedDB, and loss and per-class accuracy are charted, with thirty practise shapes built in.

Points are scattered at random across a square and labelled only inside or outside the circle. A binary classifier given that one bit of feedback and no geometric prior draws the boundary itself in WebGL — the fit visible, not described.

Two agents trained against each other, rewarded by the difficulty each causes the other rather than by winning. The page solves for the ball speed past which no paddle can intercept in time — a limit computed, not guessed.

Hill climbing rather than backpropagation: a controller is scored on survival time in a Web Worker and keeps whichever change lives longer. Shooting emerged from that reward — nobody scripted it.

Nineteen draw histories, three fitted models and a backtest against the draws themselves. No model beats chance, which is the finding; what the page can do is avoid crowd-picked numbers, and it shows the arithmetic per line.

Ten levels, three lives, and one rule that makes the fairness arithmetic: a shark cannot fire until it has turned to face you. The rules live in their own module behind three.js, with a Node server for the shared board.

Higher-or-lower in React on Ant Design components: each guess is scored against the probability of the next draw, and the best streak is kept as you play rather than reset with the page.

Six difficulties in React, with Apollo Client for data and Encrypt Storage writing games and mistakes to the device — so a half-finished board is still there next visit.

400,000 stars integrated on the GPU every frame — WebGPU compute with a WebGL 2 fallback picked at runtime — with pointer input bending the arms and a click sending a shockwave through the field.

Generated in the client with a length and character-set control, so no candidate is transmitted and there is no server that could log one. React on Ant Design; the whole tool is one page.

Frequency ranking done in the browser, then laid out with Ant Design Charts and exported as SVG or PNG — so the image is a real vector rather than a screenshot, and the text is never uploaded.

Live positions from the adsb.lol feed over OpenStreetMap tiles, with a PostgreSQL airport survey resolving a place name to an airport code; a watched airframe raises a notice when it departs, and the fence is drawn rather than described.

A globe over world-atlas geometry with every hominin find placed where it was recovered, dated to the range the paper gives, and joined by the genetic evidence for flow between lineages — each figure carrying its source.

Next.js and PostgreSQL: the original petition, the collected stories, and a wire that polls Google News for tomato coverage. Submissions pass Cloudflare Turnstile, so the petition is not a spam target.

Ten trailer builds written up as guides and served by Express from Pug templates — Helmet for headers, Gmail for contact, and nothing dynamic where nothing needed to be.

Eighty-three fishes that carry the name bass, plotted on the cladogram their genetics actually gives and drawn as SVG, because a shared shape is convergence, not kinship.

Recorded in the parent's own voice with Web Audio and MediaRecorder, stored in IndexedDB on the device, played from a board a child can use — no upload, no account, no advertising.

Next.js and PostgreSQL with ownership proved rather than claimed: the server resolves a DNS record before a listing goes live. Offers are private and submissions pass Cloudflare Turnstile.

A Python standard-library HTTP server driving a USB FM radio. The scan only reports a station when the 19 kHz stereo pilot is present, so adjacent-channel noise never appears as an entry; the spectrum is drawn on Canvas.

An MQTT broker on Aedes with LevelDB persistence and topic authorisation, shipped as a Docker image — the one piece of infrastructure here rather than a page, and the only thing on this list a server has to run.

A static index over recalls the government already publishes: searchable by brand, category and issue month, with every entry linking to the notice it came from. Node 22 build, SVG charts, no client dependencies.

Toronto food-premises inspections as the inspector wrote them: pass, conditional pass, closure, and the infractions. Built at build time into static pages with SVG charts, so it carries no runtime dependency and no query to log.

A building's own file before a lease is signed: the city's evaluation and enforcement history on one page, each claim linked to its source. Static HTML from a Node 22 pipeline, so nothing is queried at runtime.

Provincial fish advice per waterbody, with the guide's own codebook linked rather than translated — a mistranslation here would be health advice — built as static pages from a Node 22 pipeline.

Every federal corporation, its status and its filing history, published as static pages, with search running in the browser against a sharded index rather than against a server.

Ontario career-college enforcement actions — notice, order, penalty — indexed per institution and per action, built as static pages so it can be read before tuition is paid.

Employment counts by occupation and region from Statistics Canada's table, with a suppressed cell drawn as a gap: a figure that was withheld is not a zero, and the page refuses to pretend otherwise.

Consumer Price Index series by city and year, read from the published table and charted without adjustment — the number on the page is the number in the release, and the source is named.

Average hourly and weekly wages by industry and province from the published labour-force table, presented as the record it is — no ranking, no normalisation, no derived headline.

Municipal population estimates for July 1 of each year, published as static pages and kept labelled as estimates: a census count and an annual estimate are different figures and the page does not conflate them.