The Frontier Models Score Zero
Ask the best AI models about last month's vulnerabilities and they score 0.0 and 0.2. Our sovereign CVE model scores 90.86. Here is why.
READ POSTBLOG // FOUNDER'S NOTES
Founder's Notes.
Thoughts on tooling, engineering, and building companies.
-
Research -
INFRA Redis Is Not a State Sync Layer
If your real-time path is Redis pub/sub plus glue you wrote yourself, you have a science project, not a product. Lux deletes the glue.
READ POST -
Engineering How We Stopped Lying to Ourselves About Quality
Test coverage told us we were fine. Three bugs later, we rebuilt our quality system around customer journeys instead of line counts.
READ POST -
Research Eight Weeks to #1
A solo founder, one cluster, and the public MTEB leaderboard to prove it.
READ POST -
ML Engineering Vectors Are Not Ontologies
Why structural features still matter in 2026, and what spaCy plus Rust let you extract from text that embedding models alone miss.
READ POST -
RAG Your Vector Database Is a Coping Mechanism
Why every layer of your RAG stack compensates for a fundamental problem you should have solved at embed time.
READ POST -
Tooling CSSGuard: The Missing Half of CSS Tooling
Tailwind purges unused CSS. But what about HTML classes with no CSS? We built a bidirectional validator to catch orphaned classes before they break production.
READ POST