← Back to the homepage

All projects

Everything I built for the fun of it.

Twenty live experiments, almost all on their own subpath of this domain — the three from the homepage first.

/jerlang/

An Erlang runtime for the JVM

The JErlang documentation site, with its guide sidebar and the Maven dependency io.github.slayyyyer:jerlang:3.0.0.

Compiles Erlang source to BEAM and interprets it on the JVM, so OTP’s own gen_server and supervisor run unmodified — a working prototype, not production software. Built solo: the compiler, interpreter, process model, distribution and Java interop; published on Maven Central.

  • Java 21
  • BEAM
  • OTP 29

/orders-service/

A three-repo reference system

The orders-service page: a “Recorded, not live” note above three summary cards.

One orders domain, built solo across three repositories: an API with REST and GraphQL over Postgres 16, six zero-downtime migrations measured against a busy writer, and a GCP reference platform with Terraform verified against a mock provider, never applied. Recorded from one CI run, not a live demo.

  • Spring Boot 4
  • Postgres 16
  • Terraform

/buffy/

A desk companion for your coding agent

The Buffy download page, with a plush cloud mascot in a lavender panel.

Shipped for Windows, macOS and Linux: agent activity becomes an always-on-top overlay. No telemetry. Built solo: the overlay, local daemon, CLI, hooks for Claude Code, Gemini CLI, Cursor and Freebuff, and the installers.

  • Desktop app
  • CLI
  • Local daemon

/morphex/

Pluralize English nouns, deterministically

The morphex demo turning mother-in-law into mothers-in-law, with the lexicon rule that decided it.

Exact on the frozen 36,241-word set where stock lands at 96.5%. Scope: a Java 17 library — a 147,338-entry auditable lexicon, three productive spelling rules and a first-match-wins precedence, with a diagnostic API that says why a form was chosen. No dependencies, network calls or model downloads; MIT code over CC BY-SA 3.0 data.

  • Java 17
  • 147k lexicon
  • JMH

/transcend/

Transcend: English↔French from scratch

The Transcend benchmark report: 26.44 EN→FR BLEU, +2.36 over the 24.08 baseline.

A pilot-level English↔French translator: the 800M-token run improves every newstest2014 comparison against the preserved 350M checkpoint, and the training method, ablation and caveats are published. Scope: the training runs, the benchmark report, and a Windows app that runs the model on-device with no Python and no network after install.

  • 800M tokens
  • Java runtime
  • Windows app

/yahtzee/

Exact-probability Yahtzee

The Yahtzee board after a roll, dice showing 6, 3, 5, 5 and 1, with the coach recommending the fives be held.

Play against a CPU that solves every roll with exact probabilities and keeps an exact table for the rest of the game, endgame included. Scope: the exact-value solver, the table and the opponent.

  • Game
  • Exact EV

/blackjack/

Blackjack Lab

Blackjack Lab’s table mid-hand: the dealer showing 2, the player on 2♣ K♥ for 12, with a true count of +0.2 and a 0.426% house edge.

A strategy chart that re-derives itself whenever a rule changes, with card counting and a simulation target to prove the number. Scope: the exact solver and the lab around it.

  • Solver
  • Simulation

/word-chomp/

Word Chomp!

Word Chomp mid-round: a P R O S T U rack, 2:18 left on the timer and 0 of 34 garden words cleared.

Unscramble a six- or seven-letter rack, find the three-to-seven-letter words and clear every garden word before the 2:20 timer runs out. Scope: the word lists, the scoring and the round timer.

  • Word game
  • 3–7 letters
  • 2:20 rounds

/boggle/

Boggle Lab

Boggle Lab’s board — R, Y, A, Qu, E, L, F, N, C, W, O, R, O, S, T, S — beside its solver’s exact maximum of 104 points.

A 4×4 Boggle game backed by a solver that computes every word, every legal path and the board’s exact maximum score rather than sampling them — race a CPU whose risk changes near the clock, or hill-climb a high scorer in the board lab. Scope: vanilla HTML, CSS and JavaScript — no framework, no build step, no server.

  • Game
  • Exact solver
  • Vanilla JS

/earth-radio/

Earth Radio: one day of the planet, in sound

Earth Radio’s world map with 43 USGS earthquakes from November 18, 2026, the day’s peak at M6.1 and a 24-hour quake-density strip.

A single UTC day of real USGS earthquakes, magnitude 2.5 and up, becomes a 90-second ambient piece: longitude places it in the stereo field, depth sets the pitch in D minor pentatonic, and magnitude sets the loudness. Scope: a browser instrument — no samples, no backend.

  • Browser instrument
  • No samples
  • USGS feed

/clt-almanac/

The Charlotte Almanac

The Charlotte Almanac forecast card: 64° high, 35° low, 1% chance of rain.

A daily weather note for Charlotte, retrained every morning on the station record back to 1939, with a backtest that only ever sees what was known that day. Scope: ridge and logistic regressions, the backtest, and a writer handed the finished numbers with orders not to invent any of them.

  • Forecast model
  • Backtest
  • Charlotte

/murmur/

Murmur

Murmur’s calm face beside the “Calm — settled and quiet” readout, a 62°F Charlotte weather card and AQI 50.

A small animated face in Charlotte that reacts to weather, air quality, alerts, river levels and earthquakes — a simulated mood, not a real one. Scope: the character and the live feeds behind it.

  • Public sources
  • Animated face

/what-changed/

What Changed Today

What Changed Today’s November 18, 2026 summary: 536 new records, 4 changed and 0 gone across Charlotte 311 and the Federal Register.

A daily field-level diff of Charlotte 311 service requests and the Federal Register that shows only what moved, labelling records new, backfilled or aged out — and publishes a failed fetch as no snapshot rather than a reassuring zero. Scope: GitHub Actions diffs and commits, Pages serves the result; no server, database or keys.

  • Daily diff
  • Charlotte 311
  • Federal Register

/mx-lab/

MX Lab: seven seasons, measured

MX Lab’s overview: “Seven seasons, two series, measured.” beside findings of ρ 0.727 for lap 1 to finish and +27.3% Brier skill for qualifying.

Seven seasons of AMA Pro Motocross and Supercross, every moto of every round, with six analyses published on top: lap chart, fade, lap one, qualifying, titles and an Elo rating per class. Scope: the committed dataset and its build — byte-identical rebuilds, a gate that fails the build on parser errors, and warnings that name the sheet. Unofficial, derived statistics only; no source PDFs stored or served.

  • Dataset
  • Seven seasons
  • Elo ratings

/motobuzz/

MotoBuzz

MotoBuzz’s buzz meter at 50 of 100, a 24-hour chatter chart peaking at 5 records, and Monster Energy Supercross leading the headlines.

Ranks what the motocross world is talking about from public news and social sources, refreshed once an hour. Scope: the hourly collection and the ranking.

  • Ranking
  • Hourly

/saf-explorer/

SAF Explorer

SAF Explorer’s landing page above the app window open on a 2025 WinterCupOBJ.saf archive with 87 files and 89.1 MB.

Lists and extracts what is inside an MX Simulator .saf archive, or packs a folder back into one, with extraction that cannot escape the target folder and builds that write a temporary file before replacing the archive. Scope: pure Python standard library, shipped as an 11.5 MB standalone exe.

  • Python
  • .saf format
  • Standalone exe

/tadc/

The Amazing Digital Circus: a fan-made break room

The circus big top at night: six cast members on a checkered floor beneath a red-and-blue canopy, a painted EXIT sign, confetti and the row of mode buttons.

Thirty-three small games, puzzles and toys behind a working big top: a calliope, confetti, a passport that stamps each act, and a backstage lounge where the cast answers from a character GRU trained in pure Node. Scope: static HTML, CSS and JavaScript with no build step or backend — chat and the model run in the browser, Quiet and Calm modes are everywhere, and the fan project is unofficial and non-commercial.

  • Vanilla JS
  • 33 acts
  • Tiny GRU

/j/

Think in arrays: a J interpreter in TypeScript

The j.ts playground: a J session holding i. 3 4 as a 3×4 table, a sum of each row giving 6 22 38, and a mean of 2.5, beside the list of runnable examples.

A practical subset of J — no operator precedence, every value an array — evaluated by an interpreter written from scratch in TypeScript and run entirely in the browser. Scope: the array model with exact rationals and complex numbers, rank, adverbs, trains, explicit definitions, sparse storage and J’s display format, behind a REPL that survives a reload and a ten-part guide whose examples all run. Ported from the Java interpreter in the same repository; no runtime dependencies, and Vite only builds the page.

  • TypeScript
  • J subset
  • No runtime deps

/why-lang/

Drop a file, hit its URL, it runs

The Why homepage: “Drop a file. Hit its URL. It runs.” beside a site/index.why page — imports, controller and view — rendering a Latest posts list at localhost:8080.

A small page language for the JVM: a page is a .why file in a folder, and the server compiles it on request — no router, no build step, no framework, one jar and a directory. Scope: imports, controller and view sections in a single file; mustaches that evaluate real Ruby-flavoured expressions, escaped by default; SQLite with bound parameters; file:line errors with a caret and did-you-mean hints; and why check, which compiles every page and prints a page map before you serve it. A proof of concept, deliberately small.

  • Java 21 · JVM
  • Single jar
  • SQLite built in

whyx-sxzp.onrender.com

280 characters, one .why file per URL

The whyx home feed signed in as Alice Chen: a compose box above posts from Alice and Erin Walsh with reply, repost and like counts, under Home, Explore and Profile navigation.

An X clone built from Why pages on a vendored copy of the runtime: feed, threads, profiles, replies, likes, reposts, search, sign-up and sign-in, with no router and no page build step — one .why file per URL. Scope: server-side sessions with sign-in rotation, PBKDF2 password hashing and SQLite through bound parameters, with the runtime's own suite driving the real pages through a test server — including the home feed's follow graph and a check that nobody deletes someone else's post. Ships as a two-stage Docker image on Render, where the free instance rebuilds the demo database from its migrations on every wake.

  • Why pages
  • SQLite
  • Docker on Render