Notes from the workbench
I build AI systems that have to work in the real world.
I’m Gareth Austen, a software and machine-learning engineer in Dublin. This is where I write about optimising Agents, developing robust quality gates, designing good evals, building Tell Me Ted, working with coding agents, and turning prototypes into dependable products.
Latest writing
All writing →-
Why Our First AI Architecture Was Too Complicated
Why an architecture built around separate AI capabilities became too fragmented for the continuous reading experience we wanted.
-
How We Turn Product Ideas Into Codex Tasks
A repository-aware workflow for converting product ideas into Codex tasks with invariants, tests, review prompts, device validation and rollback instructions.
-
The Most Important Product Decisions Were Not Technical
The hardest Tell Me Ted decisions were not model choices. They involved the teddy, the book, the button, the wake word, latency and when not to listen.
Series in progress
Building Tell Me Ted
Twelve field notes on building a physical AI product: the architecture, the hardware, the judgement calls and the mistakes.