Notes from the workbench.
Short, technical posts about what I'm building and what I'm learning — architecture decisions, AI-assisted development lessons, experiments that worked (and ones that didn't), and product concepts before they're products. Deliberately generic and technical: how I build, not anything client- or company-specific.
Teaching an AI agent to read a 90-page PDF and cite its sources
How I built page-aware extraction with citation validation, so every answer points back to the exact page — and why "trust me" output is unshippable in compliance software.
One engine, three products: reusing a document-understanding pipeline
The shared core behind the Leo suite — and the discipline required to keep it generic instead of forking it per product.
Structured JSON as a contract between the model and the app
Why I make the model return strict JSON schemas instead of prose, how I validate and repair them, and what that buys in reliability and testability.
Eval harnesses for AI features: keeping prompts from silently regressing
Building a lightweight evaluation suite so a prompt change can't quietly break extraction quality — treating prompts like code that needs tests.
From concept to pilot in a weekend: anatomy of an AI-assisted build
A walkthrough of the agentic-coding workflow — Claude Code, Cursor, Codex, Next.js, Vercel — that takes a scoped idea to a working MVP without a team.
Turning a published regulation into software you can trust
How I encode public rules (like the FTC Funeral Rule) into deterministic checks plus AI judgment — and where to draw the line between the two so the output is defensible.
Follow the build.
I write these as the products take shape. Want them as they land? Reach out and I'll add you to the list — a proper subscribe form is on the way.