Insights & Articles

10 essays published

Reliability

Reading the document is the demo. The rest is the product.

Any model can pull fields off an invoice. What separates a document demo from an application is what happens next: which path it chose to read it, whether it knows it might be wrong, what it cost, and where the document went. Measured on one that has been public for a month.

Sep 13, 2026 · 14 min read

Data Engineering

From demo to production. The gap is not the code.

A script on a laptop and a service that runs every morning are separated by four things: a server, a container, a volume and a scheduler. Here is each one, what it cost, and the three ways the finished thing can fail without telling anybody.

Sep 13, 2026 · 12 min read

Cost Engineering

Nobody chose 4 GB. It was the default.

A job that peaked at 66 MB was billed as though it held four gigabytes for its whole life. Then the laptop overstated the peak by 2.4x, and the small test run understated it by half. Three memory numbers, all wrong, in three different directions.

Aug 26, 2026 · 9 min read

Measurement

I almost published 37,091 against 2,399. Both were noise.

A public endpoint, one sum, and a comparison that decides what you call a product. It reversed direction when I changed a parameter that does not mean what it says, and 92% of the winning number turned out to be two Google Maps scrapers.

Aug 26, 2026 · 10 min read

Systems Design

Every fact was correct. The answer was nonsense.

Four tools each returned accurate data, every field traceable to a real page. Nothing crashed, nothing was empty, the run was green, and the pipeline allocated a brand's budget to a television station's advertising desk. The failures were all in the seams.

Aug 23, 2026 · 11 min read

Pipeline Reliability

Your pipeline didn't fail. It lied.

A pipeline that fails is a nuisance. A pipeline that succeeds and returns wrong data is a different animal. Four silent failure modes from a real extraction pipeline, and what actually fixed them.

Aug 16, 2026 · 8 min read

Extraction

Your parser didn't miss the address. It made one up.

Twenty-nine passing tests, and two thirds of the addresses returned did not exist. Not malformed, not stale, but assembled out of ordinary English sentences by a pattern doing exactly what it was told.

Aug 18, 2026 · 11 min read

Systems Design

Harness and Scaffolding: The Two Structures Around Every AI System

Scaffolding is what you build because the model is weak. A harness is what you build because the model is strong. One depreciates with every release. The other compounds.

Aug 2, 2026 · 7 min read

Infrastructure

AI as Infrastructure: The New Energy of Intelligence

AI is not a tool you use. It is becoming a layer you depend on, like electricity, like the internet. And most people building on top of it do not own any of it.

Apr 22, 2026 · 7 min read

AI Engineering

Designing Predictable AI: A Framework for Balancing Consistency and Creativity

Unpredictability is not a failure of AI. It is a property of intelligence. The real problem is building systems without deciding when predictability is required and when it is not.

Apr 24, 2026 · 6 min read