Selected Architecture & Engineering Works for Preferred Networks (PFN · プリファードネットワークス)
pureQTern & project-zero
Zero-dependency 1.58-bit ternary codec in both C and Rust agreeing to the exact bit on a golden SHA-256 hash, paired with a bare-metal C inference server.
8b-is / MLX-QUANT
Custom low-bit ternary quantization and packed matrix-vector kernels eliminating memory-bandwidth bottlenecks for foundation model inference.
Headroom (PR #1418 — kompress)
Per-content-type bias multipliers for domain-aware compression routing in Headroom.
Osaurus & Osiris
Core Architect on Osaurus (native Swift agent runtime) and Osiris (real-time OSINT intelligence engine).
deepseek-go
High-performance, low-allocation Go client and streaming runtime for foundation model APIs.
ZENO (銭 · 禅 · 善)
Cryptographic double-entry Kintsugi SHA-256 ledger and high-precision settlement engine.
Why Preferred Networks (PFN · プリファードネットワークス) (Otemachi, Chiyoda-ku, Tokyo (大手町フィナンシャルシティ))
Preferred Networks stands virtually alone in Japan in owning the entire vertical stack — from custom MN-Core systolic matrix silicon and compiler toolchains down to sovereign Japanese foundation models (PLaMo) and black-box optimization (Optuna / Rustuna). My engineering obsession is that exact boundary where mathematical structure meets hardware memory layout: writing zero-dependency C and Rust ternary codecs (pureQTern) that agree to the single bit, and custom GPU kernels (MLX-QUANT) that cut memory bandwidth by 12.80×.
Preferred Networks（PFN）が推進する独自AI半導体「MN-Core」シリーズから国産基盤モデル「PLaMo」、最適化基盤「Optuna / Rustuna」に至る垂直統合技術は、世界のAI計算基盤の中でも他に類を見ないものです。C言語とRustで1ビットの狂いもなく一致する三値量子化コーデック（pureQTern）やカスタムGPUカーネル（MLX-QUANT）をゼロ依存で実装してきた私のシステム工学スキルを、大手町のPFNチームでMN-CoreランタイムおよびPLaMo推論基盤の高速化に捧げたいと強く志望しています。