LLM Evals Start as Thirty Test Cases, Not a Platform
Most LLM features ship on vibes and break on the next prompt tweak. A small regression set, cheap deterministic checks, and a judge you actu...
Nov 1, 2026
4 min
Prompt Caching: Your Prompt's Prefix Is a Cache Key
Prompt caching can cut repeated input costs by around 80%, but only if your prompt is ordered so the stable parts come first. Here's how it ...
Oct 31, 2026
4 min
The agent that refused to drift
The agent and its job I have a loop-agent running on my MacBook for about 6 months.
Oct 29, 2026
11 min
Cutting LLM API Costs: Prompt Caching, Batching and Smaller Models
An LLM bill grows faster than traffic. Here's how prompt caching, the Batch API and smaller models cut it, with real prices, tested Python c...
Oct 28, 2026
15 min
MLX: Machine Learning Built for Apple Silicon
Discover how Apple's MLX framework brings efficient machine learning to Apple silicon. Learn about unified memory, lazy evaluation, GPU acce...
Oct 22, 2026
6 min
Structured Outputs: Stop Parsing JSON Out of LLM Prose
Constrained decoding means the model can't return malformed JSON. Here's how structured outputs work with Zod in TypeScript, and the edge ca...
Oct 24, 2026
4 min
I Built the Smallest LLM I Could to Understand How They Work
I use LLMs every day without knowing how they work, so I built the smallest one I could: a tiny transformer that writes Shakespeare one char...
Oct 23, 2026
9 min
Claude Sonnet 5.5 Migration: Three Settings to Check and What Three Tasks Cost
A newer model with an unchanged price table can still raise or lower your bill per task.
Oct 22, 2026
6 min
Write an AI Agent Loop in Under 100 Lines
Frameworks make agents look complicated, but the core is small: a model, a few tools, and a loop. This post walks through a working coding a...
Oct 22, 2026
5 min