Insights on privacy engineering, local AI, and the craft of building software that respects its users.
A comprehensive look at NotesXML — 22 note types, on-device AI with 11 models, PDF annotation, voice transcription, and the Zero-Exposure Protocol that keeps your data entirely on your device.
Meet the free AI assistant that runs entirely on your device. Voice-first interaction, GPU-accelerated inference, and zero data collection. Available now for Android, Windows, and Linux.
Beyond chat — an AI that plans multi-step workflows, executes them using real tools, and delivers results across your notebooks and filesystem. All local.
1,264 tests passed, zero failures. A detailed look at the verified capabilities and automated test results for the current NotesXML release.
Six non-negotiable principles that govern how we build software — from zero-tolerance for patches to AI-assisted development with human-owned responsibility.
How a human-AI development partnership, governed by formal specifications and continuous feedback loops, produces a self-improving development methodology.
One codebase, three platforms. How NotesXML is built, tested across automated and manual layers, and hardened through Development, Beta, and Production build phases before it reaches your device.
NotesXML notebooks are stored in .nxl files — an open, XML-based, human-readable format. The complete v2.6 specification is now published under CC BY 4.0 so anyone can build readers, writers, converters, and integrations.
Complete benchmark results for the 11-model catalog tested across eight purpose-built evaluation suites on Ryzen 7 7800X3D + RTX 5070 Ti. Eight catalog models sit on the Pareto efficiency frontier; two are retained for tier-essential capabilities (licensing, Desktop-tier vision). Includes HMS ranking, per-model breakdowns, and Pareto analysis. HMS leader: Gemma 4 E2B (0.884).
Full benchmark data for every model evaluated in the 44-model sweep. Includes per-model scores across eight benchmarks, HMS rankings, and a performance-vs-speed scatter plot. Reference companion to the catalog recommendations article.
Full question bank for the 50-question HEE exam across Linguistics, History, STEM, and Health/Logic. Complete answer key and grading rubric mapping scores to educational equivalency levels.
All 200 questions across 10 licensing domains: MBE, Electrical PE, Chemical PE, EPPP, USMLE, NCLEX-RN, Real Estate Broker, CPA, ARE, and CISSP. The most professionally targeted benchmark in the suite.
100 doctoral-level questions spanning Kripke semantics, Gödel's theorems, philosophy of language, epistemology, and metaphysics. The hardest benchmark in the NotesXML suite.
100 questions generated at runtime from 10 logic templates × 10 semantic themes. No question can have appeared in training data. The most contamination-resistant benchmark in the suite.
20 tasks across 4K–128K token context tiers. Tests needle retrieval, position bias, multi-fact synthesis, and long-range reasoning. VRAM cascade detection prevents silent score inflation.
25 tasks across 6 categories: schema compliance, tool selection, parameter extraction, multi-step sequencing, error recovery, and appropriate refusal. JSON-graded only — no tools are ever executed.