A running log of what I've built, what I've tried and dropped, and what I've looked at and decided not to touch. As much for me as for anyone else. Filter by topic, or search inline.
40 days, 2,695 settled trades, real money. My best strategy won 93.2% of its bets and still finished down.
I run a 26 billion parameter model on a laptop. Only 4 billion of it works on any given word, and that unlocks a complete new world.
Five of six orders came back empty. I blamed the market. The market was innocent and my own maths was a hair too cheap.
Ollama took 126 seconds, whereas oMLX only took 1.8 seconds. How I run a free ChatGPT on my own Mac, and what it does for me all day.
My backtest handed me a Sharpe ratio of 9. That is not a result. That is a receipt for something you did wrong.
340 settled trades. One test says I beat a coin flip. The other cannot tell whether I make money.
90 trades, a 60% win rate, and a book half a trade from flat. Why I swapped open-ended forecasts for 300-second windows.
Booking tools rent you back your own data. Mine reads iCloud over CalDAV, prices bots out in layers, and attacks itself on every deploy.
Four Polymarket contracts, four deadlines, one settlement inside six minutes. The risk was never correlation. It lived in the resolution rules.
Gmail binned my newsletter while every check that proves a sender is genuine passed. The first of three failures, and every one reported success.
My local-LLM Polymarket system hit a 69% win rate, until one tweet flipped the book overnight. The model was never the problem. So I changed the game, in public.
The whole analysis brain runs on free, delayed data. I pay for exactly one feed: the live options chain an order actually touches.
I replaced a 69,000-line code tool with 4,090 lines of Rust in a day, because a stale code graph is worse than none.
Most run-phase chains optimise cost. One enforces a hard gate, because getting it wrong would ship a UI bug nobody saw.
Most jobs in a planning pipeline do not need a frontier model. Picking the right one per role saves an order of magnitude. One job earns the splurge.
Porting Retrospective Harness Optimisation into a harness with real gate outcomes. What the label-free recipe buys you, and what to swap.
Eight techniques from Cloudflare's multi-agent code reviewer, ported into the integration-verify gate of my custom agent harness. What transferred and what I had to write down myself.