Ken Barry · Cork · September 2026

What I built after leaving salaried work. Here it is.

Four working systems, built alone: e-sports prediction models that beat the bookmakers for years, an agent that writes and runs its own Java, a crypto microstructure lab, and a rover. Each has its own page with the real screens. Every number here is measured from the code and the git history.

Hong Kong Island at night with searchlights and the lit ridge of the Peak, built by AuToken in its own 3D rendererHong Kong at night, built by AuToken, the agent I use all day, from one sentence. How, and two more worlds.
5,808commits I made, Jan 2025 to 15 Sep 2026, all repos. Merges, auto-commits and Codex tasks left out; GitHub's raw count is 36,588
860klines of source in the four main repos, tests included, counted with git ls-files
35repositories on GitHub, all private for now (why: below)
4systems running today: the engine, the agent, the lab, the rover

AuToken · 3D worlds built from one sentence

Say what you want to see. It writes the world.

Each scene below is a Java program AuToken wrote from one sentence, compiled, ran and photographed in its own OpenGL renderer. No model files, no imported scenes, no textures from disk.

A packed two-tier tennis stadium from above, late afternoon“Build me a proper tennis court in there, stadium, crowd, net and all, late afternoon.”19,481 generated spectators · 3.1 million triangles Central Hong Kong at night with searchlights and the Peak behind“Give me Hong Kong at night, looking across the harbour from the Kowloon side.”the island's ridge shaped on its real summits A stone cottage with a slate roof, a green estate car on the drive, trees and a pond“Create me a very visual 3D environment, and make sure it uses textures.”every texture generated at load

See the worlds, full size, on the AuToken page →

The projects

Four systems. Click any screenshot to enlarge it; the links open each project's page.

since April 2016 · Counter-Strike models since 2017 · full-time since April 2023

Lure e-sports prediction models that beat the bookmakers, and the window I watch them through

Machine-learning models for Counter-Strike, in use since 2017, that beat the bookmakers for years until around twenty of them banned my accounts; now also Dota 2, League of Legends, Valorant, StarCraft 2, Mobile Legends and tennis. TrueSkill ratings and 1,845 engineered features on the CS map model, over ~45 data and odds feeds reconciled into one event graph, retrained and redeployed every night. lure.ie is not the product: it is the read-only window I built to check on the engine while travelling, switched on so you can look in.

Java · 5,700+ commits · 222k lines · MySQL · runs 24/7 at home

Open the Lure page →

The live event list on lure.ie: one row per canonical event with source coverage and start timeslure.ie right now: today's events, one row per canonical event, source coverage per rowA Counter-Strike match chart on lure.ie: prices from four venues minute by minute on one probability axis with the model's line and scorebot incidents along the baselineSpirit v MOUZ: four venues' prices, the model's line and every round from the scorebot on one axisA Match Winner market opened on lure.ie: best price and model probability per runner, resting depth per venue and the model's prediction curve on one probability axisOne market opened: the model's probability next to the price, depth per venue under it

since August 2025 · the one I use all day

AuToken a desktop agent that writes its own Java, runs it live and gets judged blind

A goal in plain English, typed or spoken, becomes one tool-calling session: the model looks up what it needs among about 1,800 documented methods on 63 runners, writes Java that AuToken compiles and runs inside itself, and a judge that never saw the request scores the result. Ask it for a stadium or a city at night and it writes the program that builds the world; ask for five minutes of prog rock and it writes that too. Behind it: its own 3D renderer, a music and speech stack, phone and printer bridges, and the rover. I designed and directed it, and coding agents wrote much of the code.

Java · 22,800+ commits · 516k lines · Swing/FlatLaf · LWJGL · H2 · JNA · OpenCV

Open the AuToken page →

Courtside in a packed tennis stadium, the server in the trophy position, built by AuToken in its own 3D renderer from one sentenceCourtside, from the sentence “Build me a proper tennis court in there, stadium, crowd, net and all, late afternoon.” Every spectator generated, in AuToken's own rendererThe AuToken workspace: goal tree, snippet, evidence and the 3D worldThe workspace: the goal on the left, the Java it wrote and the evidence in the middle, the 3D world on the rightAuToken's eight provider badges rendered large by the app's own badge component: Agent on OpenAI Codex CLI, LLM on OpenAI gpt-5.6-terra, Speech on OpenAI 4o-mini transcribe, Image on OpenAI gpt-image-2, Video on Google veo-2.0, Sound on Suno chirp-v3, 3D on Local Hybrid, Vision on Google plus three moreThe provider badges, drawn at high resolution by AuToken's own badge component: eight capabilities, each naming the provider and model serving it. Every badge is a dropdown

December 2024 – now

CryptoAI a multi-venue market-microstructure lab

Fourteen exchange connectors, 1.5 TB of recorded order books, lead-lag and large-move studies, a market-making simulator with playback, predictors, and a live pricing service with a WebSocket book board.

Java · 430+ commits · 51k lines · Spring Boot · 14 venues · replay-driven tests

Open the CryptoAI page →

CryptoAI live viewer with order-book depth heatmapLive viewer: order-book depth for BTCUSDT futuresMaker playback workbenchMaker playback: a strategy's quotes against recorded depthThe book service boardBook service: rolling up/down markets, live

November 2025 – now

Buggy a rover built to be measured, calibrated rather than corrected

Teensy motor firmware, a Raspberry Pi 5 Java runtime, an Android remote and a desktop console, plus a millimetre-accurate computer-vision lab that measures the robot instead of hiding its defects behind a gain.

CAD · firmware · Java · Kotlin · every constant on the robot measured in its own lab

Open the Buggy page →

Buggy 0.4 in Fusion 360, x-ray view: platforms, electronics, drive and the camera head inside the shellBuggy 0.4 in Fusion 360, x-ray: two platforms, the electronics, the drive and the pan-tilt camera head inside the shellBuggy 0.1 in Fusion 360, solid view with the camera headThe enclosure with the camera head, as designedThe rover console on its own 1024 by 600 screen, captured live: both cameras, map, telemetry with 254 hours of uptime, wheel targets and the goal buttonThe console on the rover's own screen, captured live today: both cameras, telemetry at 254 h uptime, wheel targets, goal button

Commits, 2023 – 2026

The commit graph, with the months I was employed marked on it.

GitHub's own calendars for github.com/erGOD are in the strip below, read while signed in so private work is counted. They say 36,588 contributions since January 2025. That figure is inflated: it counts every Codex cloud task up to four times (the task's commit, the pull request, the merge, a branch update) and every auto-commit from my scripts. The graph under the strip is read from the repositories themselves and shows only the commits I made: 5,808 since January 2025, 1,605 of them in agent sessions at my desk. The dashed band is my last salaried stretch, dates as on my CV.

GitHub profile contribution graph, 2023github.com/erGOD, 2023, GitHub’s own calendar: it counts merges, pull-request events and auto-commits tooGitHub profile contribution graph, 2024github.com/erGOD, 2024, GitHub’s own calendar: it counts merges, pull-request events and auto-commits tooGitHub profile contribution graph, 2025github.com/erGOD, 2025, GitHub’s own calendar: it counts merges, pull-request events and auto-commits tooGitHub profile contribution graph, 2026github.com/erGOD, 2026, GitHub’s own calendar: it counts merges, pull-request events and auto-commits too
2023372 employed · Pythia Sports, to April 2023 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 2024958 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 20251,766 Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 20264,042 Jan Feb Mar Apr May Jun Jul Aug Sep per day 0 1–4 5–14 15–39 40+ employed, dates as on the CV
repolangcommitsminesizefirstlast push
autokenJava22,6432,4392,011 MB2025-082026-09-15the agent
sportyaiJava5,7185,58748 MB2016-042026-09-06Lure engine
lure-uiJavaScript1,0631,04319 MB2016-082026-09-15the lure.ie window
crypto-aiJava4384332 MB2024-122026-09-15CryptoAI
holdem-aiJava4034025 MB2023-082024-11-14poker bot (CV)
tubyJava17617110 MB2026-082026-09-14video pipeline
buggyHTML15411238 MB2025-112026-02-03rover CAD, firmware
autoken-desktopJava2424363 MB2026-092026-09-14AuToken public cut
gobanGo2423284 KB2026-072026-08-02a Go server on EC2
ergod-commonJava85507 KB2016-072019-12-21

Top ten of 35 repositories by commits on the default branch to 15 September 2026. “Mine” is the commits I made, with merges, auto-commits and Codex tasks left out. SDK forks left out.

Why private, and why not public overnight. Every repo is a working system with exchange keys, broker sessions, provider tokens and recorded market data threaded through its history. Publishing one means screening the whole history, not the tip. A screened public export of AuToken is in progress; the rest follow one at a time. Ask and I will walk you through any of them on a call.

What GitHub's 36,588 is made of. 5,808 commits I made (1,605 of them in agent sessions at my desk), 9,416 Codex cloud tasks I wrote, reviewed and merged, 9,476 pull-request events for those tasks, 10,268 merge and revert commits from the merge button, 865 auto-commits from scripts, and 755 issues, reviews and pull requests closed unmerged. Only the first group is on the graph, where the busiest day is 148. The 2023 and 2024 rows come from the same scan, so the poker bot committed under an old address is on them; GitHub's calendar leaves it out.

Video, rendered by my own pipeline

Chess lessons and market Shorts, from data to frames.

A Java pipeline (repo tuby, since August 2026) takes a data question, runs the analysis, writes the narration from the numbers, renders the frames (Blender for the chess table, Java2D for the market boards) and packages the result. Frames from finished videos:

Bishop and knight mate lesson
Chess lesson: bishop and knight mate, with the Lichess conversion rate by rating
Blunders per 100 moves by rating
Chess lesson: blunders per 100 moves, from engine-annotated games
What each venue charges
Short: overround by venue
Said vs happened
Short: Polymarket calibration

Away from the desk

India, 17–31 January 2024: Agra, Mysore, Srirangapatna, Hampi.

Click any photo to see it full size.

Ken in front of Mysore Palace, lit up
Mysore Palace, 26 Jan
The Taj Mahal from the reflecting pool, Agra
Agra, 20 Jan
Ken on the Maharaja's elephant inside the Mysore Palace compound
Mysore Palace grounds, 29 Jan: the palace's own elephant, not a sanctuary. Two palace guards suggested a small bribe; it worked.
The Gumbaz at Srirangapatna
Srirangapatna, 28 Jan
Boulders and river at Hampi
Hampi, 31 Jan
Chamundeshwari temple, Mysore
Chamundi Hill, 25 Jan

Timeline

Dates, from the repositories and the CV.