In which the accounts are settled, honestly, including the ones that do not flatter anybody.
Where it stands, on the 19th of August 2026:
morning. Three run on isolated paper pools - a tactical satellite, an intraday burst experiment, and a monthly stock-options book opened three weeks ago.
The live book entered a position this morning at 09:45 without consulting me, marked it every three minutes, and filed its report at 15:45. I read the report on my phone.
What the AI did superbly. Implementation velocity that still surprises me - the entire options-greeks layer, correct, in an afternoon. Exhaustive parameter sweeps run without complaint or corner-cutting. Regression tests written unprompted, including one occasion where it proved its own new tests would fail against the old code before accepting they were meaningful. And a particular kind of stamina: the eleventh rejected exit rule was investigated as carefully as the first.
What the AI got confidently wrong. Margin figures misread by 40× and then 3×, in opposite directions, from two different wrong fields of the same API response. A credit calculation derived with the wrong formula, reported with a confident explanation attached. Code that re-priced a position the book had already closed, inflating unrealised profit by the exact amount that had been spent closing it. A closing routine that unwound in the order that maximised risk. Arithmetic that produced a buy limit below the ask on every protective leg the book buys.
Note the shape. Not one of these is sloppy. Every one is plausible, idiomatic, well-commented and wrong about a specific fact of the world.
What only the human could do. Deciding to stop grading my own homework and buy real market data - the single highest-leverage decision in the project, and one no amount of model capability substitutes for. Killing the flagship strategy. Deciding that a variant which won across all history but lost in the current regime does not ship. Noticing that a summary table did not add up. Asking why the concentration metric measured only the call side - a question that invalidated every historical reading of it. Asking whether using the post-auction cash close for intraday decisions "seems wrong" - it did, and a guard now suspends that logic for the final fifteen minutes of every session.
I did not write most of this code. I made every decision that mattered.
1. Define the seams before you delegate. Two days of interface design bought three months of two subsystems never disagreeing. The bottleneck is not the model's ability to write; it is whether you have decided what the pieces are.
2. Structural limits beat review. You will not catch elegant-but-wrong code by reading it. Bound the worst case by construction, hardcode the risk limits so changing one requires a commit and an explanation, and stage from simulation to small live to full size.
3. The scarce skill is spotting plausible-but-wrong at speed. This is now the job. Not writing code - auditing confident output against reality, fast, repeatedly, without becoming either credulous or paranoid. It is a genuinely different muscle from programming and I am noticeably better at it than I was in May.
4. "No change" is a result. Sixty-seven studies, mostly rejections, and the rejections produced the most valuable finding in the project. Judge research by conclusions you can trust, not experiments you can run.
5. Fund the measurement before you fund the model. Two months of conclusions rested on reconstructed prices with an invented bid/ask. Real data inverted several of them. AI makes building and testing cheap; it does nothing about what you test against.
6. Verify the instrument. One reconciliation constant that every experiment must reproduce before it may report. It costs one line and converts silent corruption into loud failure. It has already saved this project twice.
7. Be willing to kill your best work. The most sophisticated component was the one that did not earn its keep, and a rule you could write on a napkin is what trades today. AI makes elaborate machinery cheap, which makes the ratio of elaborate-to-necessary in your systems worse, not better. Measure the clever thing against a stupid baseline before you fall in love.
It is not a finished system. The margin path has an open question that needs a portfolio-level simulator neither the engine nor I have built. The server is undersized and swaps. The win/loss ratio on the live book sits at 0.73 against a stated floor of 1.0, and until that resolves the book stays at reduced size. There is a tracker item, opened three weeks ago, that says a skipped entry produces no alert - still true, still open.
Nor is it a claim that a solo operator with an AI can outperform a trading desk. It is a claim that a solo operator with an AI can build, test, deploy, monitor and honestly evaluate a production system in three months, at a standard that would previously have required a small team - and can know, with evidence, which parts of it work.
The market remains entirely indifferent to how the software was written. That is the useful thing about markets, and the reason I chose this as a proving ground rather than something more forgiving. You cannot argue with a settlement price.
Thank you for reading. Every figure in this series came from the project's own commit history, test suite and research ledger - extracted by the tool that helped write it.
This series was written with Claude, the same tool it describes. I supplied the project, the judgement calls and the arguments; Claude supplied the drafting, and dug every figure out of the repository’s own commit history so I could not flatter myself from memory.
Declaring that seems the least I can do given the subject. It would be a peculiar hypocrisy to publish eight posts on harnessing AI while implying I typed them all by hand. If the writing is good, that is partly the tool. If the judgement is sound, that part is mine. Distinguishing between those two things is, as it happens, what the entire series is about.
By day I run product for data strategy and operations at Condé Nast, where the brief is customer identity: the unglamorous business of establishing that the person reading on a phone in Mumbai and the one subscribing on a laptop in London are the same human being. Essentially the ‘slow work’ of turning unknown into known, in various stages. The glamorous parts of my day: developing the single customer view, and harnessing that data to optimise for amplified engagement and revenue across multiple lines and brands.
Twenty-four years of it now, across product, data and technology - client side and agency side, in media and publishing, CPG, insurance, automotive, FMCG and telecom, across North America, Europe and Asia. Enough time in front of CXOs to have learned that a business case travels further than an architecture diagram, and enough time behind them to know the diagram still has to be right.
This project was my evenings. It brings together the triumvirate - my love for the world of finance and markets, my drive to build a production grade system using the latest AI toolset, and the itch to discover first hand what these tools are truly capable of. And the only honest way to find out what these tools can carry is to hand them something that can lose real money.
Questions, disagreements, war stories from your own build, or a conversation about senior product leadership and AI-delivery roles - all welcome.