We publish what we can.

In-kernel mechanism for Aether application filtering. Hashes and verdicts only; the policy lives in userspace.

Device-side security service for Aether: attacker-reputation enforcement and application-level visibility at the access point.

TR-369 / USP 1.3 agent daemon for OpenWrt access points. Full UCI backend, Wi-Fi 7 modes, hybrid X25519 + ML-KEM-768 TLS. Reports to Aether or the standalone OptimACS.

The OpenWrt package feed that carries ac-client and our device packages for MediaTek Filogic and Qualcomm IPQ targets.

Helm charts for OptimACS, the self-hostable TR-369 / USP access-point management edition.

CUDA backend for mistral.rs built on pure Rust GPU kernels instead of a vendored C++ dependency. The inference path behind our private LLM gateway and the engine behind titan-engine.

Desktop browser with an AI agent whose prompt-injection defence is architectural: a privileged planner that never reads page text, a quarantined reader with no tools, an action judge that can only escalate, and a local classifier that withholds flagged chunks. 250 tests, 90 end to end, with the models scripted to be compromised.

NVlabs cuda-oxide plus typed PTX intrinsics for cuda_device::math — the toolchain the kernels are written against.

Hand-written Rust GPU kernels for quantised LLM inference: a no_std nvptx64 crate with the f16 conversion paths rustc does not expose. The GPU path under the alive runtime; measurements and retractions are in the repo.