-
Understanding a Smart-Contract Honeypot Detector: Circuits, Redundancy, and Causal Testing
June 28, 2026
I’d like to share a small experiment on a question that comes up a lot once you start using machine learning for security: when a detector flags something, how do...
-
measure: turning "the LLM seems fine" into a number
May 20, 2026
A surprising number of LLM applications ship without any quantitative evaluation at all. Someone tries a handful of prompts, the answers look reasonable, and that’s the bar. The problem is...
-
guard: catching secrets and prompt injection before they ship
May 15, 2026
I built guard after noticing how often the same two mistakes show up in AI codebases: someone commits a real API key by accident, and someone wires up an agent...
-
router: most queries don't need your biggest model
April 10, 2026
The default pattern I kept seeing in small LLM apps is: pick one model, send every query to it. That’s simple, and it’s also wasteful, because most queries in a...
-
retrieval: an answer is only as trustworthy as its citation
March 22, 2026
Ask a plain language model about your company’s internal policy or a document it’s never seen, and it will often answer anyway, fluently and wrong. This isn’t a bug you...