Evals with a model as judge
Scoring prsist’s note-taking against sessions I labelled by hand, with Claude Opus as the judge, separate dev and test splits, and three repeats per run so I can see the noise.
Getting two model families to review each other
In cstack, Claude and Codex try to break each change, and a model from the other family settles disputes. They catch different things.
Running agents on an always-on box
Tailscale, a shared account hub and T3 Code, so I can start work from my phone and pick it up on my laptop.
Putting a price on every model call
Reply Desk counts the cost of every read, score and draft, and stops for the day before it overspends. Free rules filter first; the model only sees what’s left.
Keeping the model inside the rules
In KM Zero, a deterministic engine builds every option and Claude only chooses and explains. If the model fails, the runner still gets a correct plan.
MCP servers and hooks
Wiring one memory service into Claude Code, Codex and the Claude apps: hooks for the coding tools, an MCP server for the chat apps.
Writing task briefs agents can finish
For MRR Clock I planned the app as small, test-first briefs. The agent got further when each brief named the test that proves it’s done.
Test-driven development
Built a FastAPI service test-first in Docker. Slower to start, much faster to change.
My first NLP model, at work
As a contractor at SquareTrade I built a model to classify insurance claims from their text. It reached 80% accuracy, and taught me how much the labels matter.
Courses
Supervised Machine Learning: Regression and Classification
Andrew Ng’s Coursera course. It got me into neural networks, deep learning and NLP, and was the start of the path to building with LLMs.