Field notes in agentic engineeringἀκρίβεια — exactness

One mind. Many hands. Zero slop.

I’m an AI/ML engineer and tech lead — ten years of real-world software, now built with AI agents. I publish the whole process: the specs, the adversarial reviews, and the failures. The good, the bad and the ugly of agentic engineering.

Read the build logs The good, the bad & the ugly

Field notes

The good, the bad & the ugly

The Good

Where agents genuinely deliver. Adversarial review that caught a bug I would have shipped. Contract-first codegen that hasn’t drifted once across 12 releases.

Read these entries →

The Bad

Where they burn hours. Confidently wrong fixes, flaky tool loops, review theatre — and what I stopped doing because it never paid off.

Read these entries →

The Ugly

The failures worth publishing. The licensing wall that killed a business model after the product already worked. The backtest that lied to me.

Read these entries →

Plates

Selected work

Plate I

QuantScreen

My quant equity screener — point-in-time discipline, 503 tickers, ~50 KPIs. Every PR planned by spec and adversarially reviewed; the full build is the flagship case study.

Live · the case study

Plate II

Mage Survivor

A bullet-heaven incremental RPG I’m building the same way — different domain, same discipline, same public record.

In development

Plate III

The Newsletter

AI & Tech, distilled from many sources into one considered digest — what I’d want a colleague to forward me.

Ongoing

Contents

Build logs

Adversarial review, or: how three agents killed my favourite bugIn production The licensing wall — validate the business before you build itIn production One schema, two languages: contract-first codegen without driftIn production