Privacy

Accept optional first-party analytics or decline. Functional journey and sound preferences stay on this device.

Read the privacy notice

Outcomes

Projects show what I built. This page shows what changed because of it, in plain terms first, with the number behind each one and a link to how it was measured.

Faster for the people using it

Heaviest assistant data loads

One caching pass cut the slowest Alex OS load by 83%. The admin went from waiting to working.

JavaScript on first visit

The homepage ships 27% less JavaScript, so it becomes usable sooner on a phone.

Render-blocking CSS

116 kB of styles for three retired homepages were still shipping. Removing them means the first paint waits on less.

Keeps working when parts fail

Answers with the AI switched off

When the model provider is down or out of credit, the assistant still gives a cited answer. That was tested, not assumed.

Blank errors in the failure plan

Every dependency (vector search, embeddings, model provider, database) has a named fallback a visitor actually sees.

AI quality you can check

Right answer in the top three

Measured on a hand-labeled question set, published with its method and its limits.

Overturned by my own benchmark

The retrieval setup I shipped lost to a simpler one. The case study now says so.

Cost per answer, traced

Every model call records tokens and cost, so AI spend is a number on a dashboard, not a surprise on an invoice.

Easier to change next year

Templates on one design source

Every page and detail template reads the same Theme DNA, so restyling the site is one edit, not 23.

Content types

Consolidated without losing content, so there is less to maintain and less for the assistant to confuse.

Automated tests on Alex OS

Changes get checked by the suite before they ship, not by visitors after.

How these were measured

  • Assistant quality, latency and cost: the Alex OS evaluation report, with the dataset and scripts described.
  • Page weight: Next.js build output and the CSS bundle before and after, from the performance passes in August 2026.
  • Load times: timed on the admin endpoints before and after the read-through cache.

The evidence behind the numbers

Dataset, baselines, what lost, and what changed because of it.

Evaluation report