AI shouldn't live in a datacenter.
Altrusian Computer is building the runtime, protocols, and tooling for distributed, on-device intelligence. Your phone, your laptop, the MacBook on the next desk — that's the substrate. We treat the network as the computer, not the place compute hides.
IronContext
The Snyk for AI agents — a Rust-core security & optimization engine for the Model Context Protocol. Static analysis on every tool description the model sees before it reasons.
Ten CVE detectors for the May-2026 MCP threat pack (tool-poisoning, invisible Unicode, cross-tool shadow, rug-pull, confused-deputy, …). A Reasoning-Impact Score that quantifies hallucination risk per tool. A pluggable description optimizer that reclaims context-window tokens — all behind a single static binary and a zero-dep Python wrapper.
Naina
An embeddable C++ runtime for face & person understanding — detection, alignment, recognition, liveness, tracking, and re-identification — through one stable C ABI with Python and Node bindings.
Ship the same model to a Raspberry Pi, a phone, an Apple Silicon laptop, or a CUDA server with no code change. Backends auto-select at runtime — ONNX Runtime, NCNN, OpenVINO, CoreML, TensorRT, ExecuTorch.
See on GitHubMosaic
An open curriculum on the systems behind modern AI — from the C++ memory model up to MLA, FlashAttention-3, on-device inference, and a LAN of phones running 70B.
88 lessons across seven tracks. Free, no signup, modular fifteen-minute formats. Built because we couldn't find what we wanted to read.
Read Mosaicscree + breccia
Two open ML primitives in the same geological lineage — scree for variable-length tensors, breccia for block-scaled low-precision. NumPy / PyTorch / MLX / JAX backends, Triton fast paths.
scree ships a typed values+offsets array with FA-2-style Triton forward and backward — 1.30× of FlashAttention-2 forward, 1.61× full training step on H100, 71–85% memory saved vs HuggingFace padded. breccia is the block-scaled FP8 / FP4 / INT4 sibling, with Triton scaled-matmul at FP32-parity correctness. A breccia ScaledTensor wrapped in a scree.Array is the canonical low-precision varlen batch.
Streaming LLM UI Kit
Every AI chat re-renders the whole conversation on every token. We open-sourced two libraries that don't — StreamMD for incremental markdown, ZeroJitter for canvas-based text streams.
StreamMD turns 500 re-renders into ~20 by tracking block boundaries and freezing finished blocks. ZeroJitter eliminates layout thrashing by measuring off the main thread and painting to canvas — 0 reflows, 60fps, <1ms layout time.
Teleport
Move files and clipboard between any two devices on the planet, browser to browser. No uploads. No accounts. No third party standing in the middle of your data — perfect for moving model weights to the edge devices that will run them.
Built on WebRTC. The signaling layer is the only thing we touch, and we throw the room key away the moment you close the tab. End-to-end encrypted by construction.
Open TeleportThings we built for ourselves and shipped because they were already there. Not the studio's mission, still useful.