Urai - Unreasonable effectiveness of code-mode agents

Structured tool calling stalls on multi-step work. Given a sandboxed JavaScript runtime instead, Gemini 3.8 Flash scraped a website, read a knowledge base and built a print-ready slide deck in five minutes for fifty cents.

jevlite - Experience report in training the Decision Model

Building jevlite's training set from Super-NaturalInstructions and synthetic support and SIEM tickets, the teacher-labelling bug that cost 13 hours, and the QLoRA run on Gemma 4.

jevlite - A decision model based on LLMs

An open source take on TypeSafe's Jev. jevlite scores a fixed set of options from a single forward pass over the logits, so it returns real probabilities and cannot invent an answer.

literstream - Rust port of LiteStream

How a Rust port of Litestream streams SQLite's write-ahead log to object storage for point-in-time recovery, and why idle databases cost almost nothing to replicate.

Why I built 🤦 slop — My coding agent

Every AI engineer should build their own coding agent. The choices behind slop: the agent loop, manual context handoffs, tool granularity, sandboxes and tracking token spend.

Wiring against the vibe

Agents made typing cheap, but typing was never the bottleneck. What Accelerate, Kill It With Fire and Wiring the Winning Organization say about running teams in the AI era.

Reshaping data with PostgreSQL pivot tables

What does a CTO do?

Don’t Rewrite, Refactor: Why Brownfield Development Works

My thoughts on Substrate

Why we refuse fixed bid projects

My thoughts on Flutter

How we hire engineers

My views on the CKA exam

Tarka Labs and Agility

SSH Recipes in Go — An interlude