ai infrastructure
The stack we build to train on.
A self-hosting math-native compiler, an OS for real boards, LLM inference running on our own language, agent runtimes, and a from-scratch GPU training stack — all in development.
MAX
A self-hosting, math-native language: Unicode operators and a byte-identical self-compile, built for neural workloads, OS work and GPU code.
max-chain
A permissioned proof-of-authority ledger written in MAX, with consensus, ledger and contract logic in one language.
LLM Inference
Language models on MAX — Qwen3, Falcon-H1, and a clean-room MAX port of an 80B hybrid model that reached 93% of a reference engine's decode speed in our tests.
MAX OS
An operating system written in MAX, designed to test research-derived kernel policies against classical ones and to run bare-metal on real boards.
KeiSeiKit
A model-agnostic coding-agent stack — a native TUI cockpit, a multi-provider agent daemon, a content-addressed cache under every capability call, and a three-phase sleep layer where agents keep improving between sessions.
KeiSeiSDK
Standalone vertical-stack agent runtime + CLI for self-hosted LLM deployments.
GPU LLM
From-scratch GPU forward and backward kernels for Qwen3-32B; in our tests the layer backward matched a CPU reference (rel-L2 < 5e-6).
Operating discipline
A local NVIDIA GB10 for development and inference, rented NVIDIA datacenter GPUs for fine-tuning runs. Every cloud launch goes through a pre-flight memory and cost gate before the meter starts. No unattended pods.
