Available for contract work

AI systems that survive production

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. I reproduce it, fix it, and hand back a test that fails without the fix — fixed price, agreed before I start.

Fix one bug — $250, fixed price
Reproduced, patched, regression test, handed over · half up front

What actually breaks

Not model quality. The plumbing around it — the same failure classes as any distributed system, wearing an AI hat.

CRITICAL
Silent regressionA prompt or model change quietly degrades output and nobody knows until users complain. Now the dominant cause of production LLM incidents.
CRITICAL
Retry double-effectsThe same request arrives twice and something happens twice — a charge, an email, a credit deducted.
CRITICAL
Uncapped spendNo ceiling per user or per session. One loop, one bad actor, and the bill is four figures by morning.
HIGH
No deadline on provider callsThe model takes 60 seconds, the worker is tied up, and under load the pool exhausts and everything stalls.
HIGH
No fallbackProvider rate-limits or goes down — and it does — and the feature simply breaks instead of degrading.
HIGH
Unbounded agent loopsTool calls tool, model re-plans, forever. Cost and latency both without a ceiling.
HIGH
Half-written stateProcess dies mid-operation. Paid but not fulfilled, charged but not delivered, stuck with no way to recover.
HIGH
Truncated output saved as completeConnection drops mid-stream and the partial answer is stored as if it finished.

How it works

  1. One bug — $250, fixed priceYou name a failure. I reproduce it, patch it, and hand back a regression test that fails without the fix plus a written account of what was wrong. Half up front. If I can't reproduce it, you pay nothing and keep the write-up of what I ruled out.
  2. Sprint — $1,500–3,000When one bug turns out to be a class of them. Scope written down before I start; everything ships with tests.
  3. Stay — fractional, ongoingFor teams shipping AI features faster than they can make them reliable. I own that part so your engineers don't context-switch into it.

Proof, not adjectives

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.
Case study

How 78 tests missed a fencing bug

The failure, why the test suite couldn't see it, the fix, and the general lesson for any lock or lease.

Full write-up →
Open source

stripe-fulfilment-kit

Fenced, leased order claims that survive retries, crashes and concurrent webhook deliveries. Cross-process atomic.

59 tests · offline · MIT →
Open source

ai-review-gate

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.

Guardrails for agents · MIT →
More

github.com/GG5533

Everything else, including the quant work.

Profile →

The checklist I audit against

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.

Free resource

10 ways your AI feature breaks in production

The full audit method, given away. Silent regression, retry double-effects, uncapped spend, missing deadlines, unbounded agent loops, half-written state.

Read it →

Worth knowing before you email

If I don't think you have a problem worth paying to fix, I'll tell you and you'll have spent nothing. The audit is deliberately cheap because it should be an easy decision — and because an audit that finds nothing is still a useful answer.

Fix one bug — $250, fixed price
Or just describe what's breaking. samihabbal5@icloud.com