Andrae Williams

I read risk signals for a living — fraud, payments, merchant behavior — before I started building the software that acts on them. These days that means full-stack apps, AI agents, and the local infrastructure to run them without handing everything to somebody else's cloud.

What I do

Risk analysis taught me to distrust a system until it has been tested against real failure. I bring that into software: building agents and tools that have to hold up under adversarial or messy conditions, not just the happy path — and running as much of it as I can on hardware I control, in a homelab built for local inference.

How I work

A few things I try to hold to, in order:

• Understand constraints before proposing a solution.
• Build practical systems, not proofs of concept that stop at the demo.
• Inspect failure modes on purpose, not just the success path.
• Prefer evidence to assumptions — including about my own work.
• Make important decisions reversible.
• Document what broke, not just what shipped.
• Use AI as an implementation multiplier, not a substitute for expertise.

Why this site exists

This is where I work things out in public — organised by topic, with the occasional multi-part series, a lab for live experiments, a work page for what actually shipped, and a reference library for what I actually reach for. There is no publishing schedule; older entries get updated rather than replaced.

Provenance

Projects and lab entries record what was built by AI and which model wrote it. Posts are mine. That distinction is deliberate and it is visible on every entry that has one.