About
Reliability is the product
A language model is a component with a failure mode, not a finished system. Most of the engineering worth doing is in what surrounds it: what gets retrieved, what the model is permitted to assert, what happens when it is wrong, and how you would know.
How I work
01
Determinism around a nondeterministic core
The model is nondeterministic; the system around it must not be. Schema validated outputs, temperature zero where an answer must be auditable, retry budgets, and a verification pass, so malformed or drifting output never reaches the next stage. If the same question and the same corpus can give two different answers, the system cannot be audited.
02
Grounding is verified, not requested
Asking a model in a prompt to cite its sources is a hope. Checking afterwards that every citation points at a passage the model was actually given is a guarantee. When grounding cannot be verified, the right output is no output. Serving a shaky answer with a disclaimer attached only teaches people to skip disclaimers.
03
Measured, not asserted
Thresholds get chosen by sweeping labelled data and reading the separation, not by picking a number that sounds confident. Claims come with the test that produced them and the conditions it ran under. When something has not been measured, I say that instead of rounding it up to a claim.
04
Cost where it counts
Route cheap calls to cheap models, cache byte identical context, budget tokens deliberately, so frontier spend goes only where frontier reasoning is genuinely required. Most prompts send far more than the model needs, and that is a solvable engineering problem rather than a fact of life.
05
Pipelines you can leave running
A batch job is finished when it is idempotent, incremental, gated on data quality, and observable enough that a failure is loud. Anything less is a job somebody has to babysit, and babysitting does not scale.
06
Ownership end to end
I run my own Linux infrastructure with zero downtime, auto rollback deploys, and I own problems from architecture through production. Everything linked from this site is deployed and operated by me, which is also why the limits pages are specific.
Background
Ten years of backend engineering before the current focus: hazardous materials management for the U.S. Air Force in C# and .NET under a Secret clearance, healthcare interoperability pipelines at Humana in HL7 and FHIR, aircraft parts systems in Java and Spring Boot, financial data aggregation, and locomotive optimisation at Norfolk Southern.
The through line is regulated environments where being wrong has consequences and somebody eventually audits the trail. That is the same instinct behind the AI work: an answer you cannot trace is not an answer yet.
I also keep independent research on geometric data representation at dimensionalprogramming.com. It is a separate interest and none of the engineering above depends on it.