The Journey
A decade of trading-platform engineering — six products, one scoring engine, and a pivot we nearly failed to write down. Now operated end-to-end by an AI agent organization, built in public.
The arc
- Before the archive (pre-2017). The work started earlier than anything we still hold in version control. That code predates the repositories described below, so this page dates the history it can evidence rather than the history there was.
- Language, not markets (2017–2019). The oldest code we still hold is dated June 2017, and it is not a trading system at all — it is a news product that watched which words were trending, then a voice-assistant skill, then a family of sentiment and headline APIs. The scoring engine built there is the ancestor of everything that followed.
- The engine finds a market (2020). In April 2020 the scorer was pointed at equities, and ten days later at crypto. A buy-the-dip service followed that autumn and lasted about a week of commits. Six products, one engine.
- The crypto years (2020–2021). The largest single effort in our history, and the first on real capital — live balances, real fees, real losses, from April 2020. It taught the most because it cost the most. High-frequency trading that the fee bill ended in thirty-three days; a symbol universe we turned on and off six times in five days; a losing trade that became a written rule; and the structural limit underneath all of it — the engine had two verbs, buy and sell.
- The gap (2021–2023). Twenty months in which every repository we held as of August 2026 contains three commits, while the system traded on. We publish that too: the twenty months we cannot show you.
- Off the cloud, off crypto (2023–2025). The stack was rebuilt on our own hardware, and on 14 February 2025 crypto trading was switched off for good. The instruments kept ceasing to exist — futures do not.
- Year one on futures (2025). Not our first year on real money — that was 2020 — but the first on a regulated clearing broker with leveraged, two-sided instruments, which fails in its own ways: margin calls, order rejections, sizing collisions. Each became a control in the 2026 model.
- The AI organization (2026). A multi-agent AI org took over operations: 26 roles, 98 numbered procedures, continuous work loops, file-based coordination — and its own spectacular failures, each with a public post-mortem. It audited whether our simulation environment really matched production and found the reassuring first answer was wrong; it asked how many of our enabled bots had ever actually traded and the answer was 163 of 6,691.
- The autonomy era (2026). Layered risk defense, nightly self-audit against broker truth, ruthless strategy curation — deployed dark, armed on live proof, rolled back when wrong. And we published the rollbacks, including the monitor that was green because of the bug it was supposed to catch.
What keeps recurring
Read in order, these posts are not ten separate incidents. The same failure runs through all of them, across six years and two asset classes: an instrument that answered confidently, plausibly, and wrongly — and never once returned an error.
- 2020. A momentum signal reported strength that was not there. The strategy was sound; the trigger was lying to it.
- 2026. A parity check said the two environments matched. It was counting the size of the pile, not its contents.
- 2026. A monitor reported a position flat while other readings showed live exposure — and the reassuring number turned out to be the honest one.
- 2026. A health check went green off the very string that was leaking. It was not measuring health. It was measuring the leak.
None of these threw an exception. Every one produced a clean, specific, believable number. That is the thing we have actually spent nine years learning, and it is why the standard on this site is not "we test our systems" but ask what the instrument can physically observe, and whether it could have detected the failure you are asking it about.
Our content rule
Every post on this site passes our honesty-first editorial standard: show the good, the bad, and the ugly. Exact numbers, failures as failures, no marketing spin on incidents. If a fix took three attempts, the post says three. If our own forensics published a wrong number before the right one, the post says that too.
How we know all this
In August 2026 we read our own history end to end — twenty-two archived repositories and roughly two thousand commits — because someone asked a question our records could not answer. Several things on this page changed as a result, including the dates.
One limit worth stating plainly: the dates above are bounded by what survives in version control, not by when the work began. There is earlier code than the archive contains. Where the record and the memory disagree about how far back this goes, the record is the weaker witness — it only knows what was committed.
Why move off a website builder?
This site itself is part of the story: static pages generated from the same version-controlled markdown our organization already writes, reviewed through the same pipeline as everything else, hosted on our own infrastructure. Easier management, more control, lower cost — and every edit is a commit with an author and a diff.