4 min read
How I built an LLM eval harness that took AI output from 24% to 97% valid
A blind pairwise judge, a pinned prompt registry and one harness bug that silently inverted our model rankings. A practical guide to LLM evals for code generation.
Blog
Notes on AI engineering, full-stack architecture and the products I build, from LLM pipelines and evals to shipping on web and mobile.
3 min read
Sandboxing AI-generated React code: four layers that make it safe to run
When a model writes the React your users see, treat its output as hostile. Here's the defence in depth we use: a shared contract, AST validators, a quality gate and a cross-origin sandbox.
3 min read
Building a production MCP server with OAuth 2.1: lessons from shipping one
What it took to let customers connect Claude and ChatGPT to their own data safely: OAuth 2.1 with PKCE, dynamic client registration, scoped tools, live permissions and rate limits per connection.
3 min read
Spec-driven development with Claude Code: shipping a marketplace across three repos
How a knowledge base, a numbered decision log and as-built notes kept an AI coding agent accurate while building BrownBox's API, web portals and mobile app, with 858 tests to keep it honest.
2 min read
Making a whole browser game in one HTML file: procedural art and Web Audio music
How Danphe Dash, a game I built with AI, draws four Nepali landscapes in code, synthesizes folk-inspired music with the Web Audio API and holds 60 fps on phones.
