An AI lab that actually ships

AI you don't have totake on faith.

We build AI for the high-stakes moments - untrusted code, grading at scale, agents on the loose - and back every one with proof you can check, not promises. Rustbox is live, and it's just the start.

Open Rustbox

Products & Team

Trusted by teams at

The products

Secure executionLive

Rustbox

Run untrusted code without trusting it.

Kernel-enforced sandboxing on Linux namespaces and cgroups. Single binary, no Docker daemon.

8
languages: 147
adversarial exploits: 0
sandbox escapes

Open Rustbox

Software learningComing soon

BooleanStack

Serious practice that rewards depth, not streaks.

Interface creationComing soon

Zen

Describe a screen. Get a usable one.

Open and under the hood

We wrote the hard parts ourselves.

GraphStore and Hyperstack are open source. Orka compresses model weights toward 1 bit. Read the code, run the tests.

Agent memory Open source

GraphStore

Agent memory that recalls by meaning, not keywords.

Six engines behind one readable DSL: columnar numpy storage, HNSW vector search, and a markdown vault built for 24/7 agents.

engines

6

semantic recall

~127µs

library

Python

View on GitHub
Engineering harness Open source

Hyperstack

Stop AI agents from shipping plausible-but-wrong code.

Real-time ground truth for React, Next.js, Motion, React Flow, Echo, Go, Rust, and design tokens, enforced by adversarial gates.

skill gates

21

stacks covered

10+

View on GitHub
Model compression Research

Orka

Language-model weights, compressed toward 1 bit.

Learned codebooks and a hand-written GGML kernel keep reconstruction byte-exact.

per-weight target

~1 bit

Have a hard problem worth shipping?

We take on selective research partnerships when there is a path to a real system.

Get in touch