i built a cad editor in the browser, then taught an llm to use it
parsing autocad dwg, reconstructing walls and rooms from anonymous line segments, building a canvas cad editor from scratch, and handing the whole thing to claude as tools it can call
Notes on shipping production backend and AI systems. LLM evaluation, correctness-critical logic, document pipelines, and turning regulation into testable code.
parsing autocad dwg, reconstructing walls and rooms from anonymous line segments, building a canvas cad editor from scratch, and handing the whole thing to claude as tools it can call
Every SSH connection proves who the user is before your app runs a byte. I built a small Go framework on that, and here is what treating the key fingerprint as an account makes possible.
A year of running a document processing pipeline at scale through tender packs, permit applications, site surveys, and A0 drawings. The PDFs themselves rank third on the list of things that break.
How I split one search engine into three systems in Rust: Postgres for metadata, tantivy for the index, object storage for durability. The ordering a create request takes, and the quota race it avoids.
Most LLM features work in the demo and break in the real world. Here is the evaluation layer that keeps them honest — and how to measure 'correct' before you ship.
How to take a dense statute from primary source to an audit-ready detector — separating literal text from enforcement practice, and deriving tests from the regulation itself.
GDPR is usually treated as a legal checkbox bolted on at the end. The decisions that actually carry legal weight are architectural — and they are cheap early, expensive late.
The EU AI Act is written for legal teams. Here is the translation into system requirements — risk tiers, the obligations that touch your code, and what to build before the deadlines bite.
The FATF Travel Rule reads simply and implements anything but. The places literal text and enforcement practice diverge — and how to build a transmission layer that survives an audit.
When you can't see the reference corpus, you derive the tests from the regulation itself. How property-based testing makes a rules engine trustworthy on inputs you never hand-wrote.