Your AI feature works in testing. Then real traffic arrives and it
retries a request and charges twice, hangs with no
deadline, loops without a ceiling, or quietly runs
up a bill nobody capped.
Send me one of those. Within 48 hours you get a written diagnosis: what is
happening, why, what it costs you, and what the fix involves. Then the fix,
quoted after the diagnosis, with a test that fails without it.
Not model quality. The plumbing around it — the same failure classes as any distributed system, wearing an AI hat.
Everything below is public and readable. Judge the work rather than the claims.
The bug I'm most known for is one I wrote myself — and caught before it shipped. A lease that passed 78 tests and was still wrong: a provider call outliving its two-minute lease would have written its failure over a completed order, hiding a paid-for product and re-opening the order for a third charge. Every test settled an order from its current owner, so none covered a write from a superseded one. Found in cross-model review, reproduced in a test, fixed before it ever ran in production.
A handler that is idempotent on paper buys twice when the process dies between the provider returning and the record being written. The window is milliseconds wide. The reproduction, the two-part fix, and five checks.
Stripe does not promise to deliver events in order. A handler that decides from the event's type ships goods for payments that have not settled, and ships them again when the payment does. Reproduced, fixed, and five checks for your own handler.
How a Stripe fulfilment handler passed 78 tests and would still have charged one customer three times. The bug, why the tests missed it, the fix, and the general lesson for any lock or lease.
Fires the failures that only show up on run number two at your webhook handler — duplicate deliveries, crashes before and after the purchase, a late write after a lease expired, out-of-order events — and reports what it actually did. Emits a regression test you keep.
A panicking task bypasses the new SIGTERM handover, costing the exit code, the drain window and teardown. An outside contributor re-derived it independently and wrote that it held. I retracted two of my own claims in the same thread when measurement said I was wrong.
Narrowed a reported bug to the one middleware that still blocks stream completion and showed the other two no longer reproduce on main. The project's collaborator reached the same split three days later.
The failure, why the test suite couldn't see it, the fix, and the general lesson for any lock or lease.
Fenced, leased order claims that survive retries, crashes and concurrent webhook deliveries. Cross-process atomic.
Two models review a coding agent's output and block the turn if it's wrong, forcing a correction. Includes an honest record of the eight real errors it caught.
Their path-containment guard had no test; disabling it left every existing test passing. The new test is the only one that fails when the guard is removed. Merged into master by the project owner.
Everything else, including the quant work.
Free, no signup, no email capture. Ten failure modes with how to check each one in your own codebase. Most people won't run it — but you'll know whether you have a problem before you talk to anyone about it.
The full audit method, given away. Silent regression, retry double-effects, uncapped spend, missing deadlines, unbounded agent loops, half-written state.
If I don't think you have a problem worth paying to fix, I'll tell you and you'll have spent nothing. The diagnosis comes first and is priced as the two-to-three-hour job it is, because it should be an easy decision — and because you should be able to judge the work on one small piece of it before trusting me with anything larger. If I can't reproduce your bug, I refund the deposit and send you what I ruled out.
Written diagnosis in 48 hours — $250