ai infrastructure
The stack we build to train on.
Self-hosting math-native compiler that is also a graph and a blockchain, derivation-first OS proven on real silicon, LLM inference running on our own language, open-source agent runtimes, and a validated GPU training stack.
MAX
One self-hosting language that is simultaneously a compiler, a constructor, a graph, and a blockchain. Unicode operators, byte-identical self-compile, 982 machine-checked theorems, a GEMM proven bit-exact through the CUDA-target compiler.
max-chain
A permissioned proof-of-authority blockchain that is not built on top of a language — the language IS the chain. MAX-native consensus, MAX-native compiler.
LLM Inference
Qwen3 and Falcon-H1 serving natively on MAX; Qwen3.5-35B loading from a real GGUF; an 80B model measured at 66 tok/s; a clean-room MAX port of a reference serving engine measured at ~64 tok/s on an 80B model.
MAX OS
Operating system written in MAX with page-replace and scheduling policies derived from lab research. Boots in QEMU, and the same driver-as-matrix design runs on real silicon across four boards — Pi, Orin, K230, Orange Pi R2S — plus a Zynq-7020 FPGA target.
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 + full GPU backward for Qwen3-32B. Parity-validated against CPU reference (rel-L2 < 5e-6). Validated on datacenter and local GPUs, vendor-neutral.
Operating discipline
Cloud and local compute in parallel, vendor-neutral — the same kernels run across datacenter and local GPU classes with no vendor lock. Every launch goes through a pre-flight memory + cost gate before the meter starts. No fire-and-forget pods.