Durable execution
The agent resumes, retries, and finishes. A dropped model call does not lose the work.
Anyone can vibe code an agent. Few can make it survive production.
AI made software easy to build and hard to run. I productionalise agents that fail silently, burn money, and break under real traffic. Durable, observable, cost-capped.
The demo took a weekend. Production is taking months.
Vibecoding got you a working prototype in hours. Now your agent returns wrong answers and nothing flags it. The bill climbs and nobody knows why. The gap between demo and production is where projects die.
What production-safe means.
Evaluation
A test set catches wrong answers before your customers do.
Cost and latency caps
Hard limits on spend and response time. The agent cannot burn the budget.
I don't build demos.
I work on live agents with real traffic. If you need a demo or a chatbot, I'll refer you to someone else.
What I take
- Production failures on live agents
- Silent wrong answers
- Cost overruns
- Missing or weak evals
- Reliability work on systems already in production
What I turn away
- Demos
- Proofs of concept
- Copilots
- Model-selection theatre
- Chatbots
Need a demo? I'll refer you to someone else.
Engagements open with a failure autopsy.
We start with the logs.
-
Autopsy
We look at what broke in production and what the eval missed.
-
Fix
Durable execution, real evals, and hard caps on spend and latency.
-
Prove
You get a sanitised postmortem you can share with your team.
Incident reconstructions.
I take a public AI failure, rebuild the assumptions that caused it, and show exactly where production broke them. Each one covers what broke, what the eval missed, and what would have caught it sooner.
The first reconstruction is in progress.
View all writingYour agent is failing. You cannot say why.
Tell me what broke. I'll tell you if I can fix it.
Start a conversation