I specialize in ternary models and designed GF-T — a ternary-native number format that beats comparable formats on benchmarks.

I take that format from an arXiv paper to silicon — and make the neural network train itself on the FPGA. Math → spec → RTL → silicon → on-device ML, solo, on a fully open-source flow.

Open to remote contract work · Ko Samui, Thailand 🌴 · UTC+7

🎯 Looking for: hardware-AI · ML-systems · FPGA / RTL roles — remote, full-time or contract 🔨 Currently: proving on-device neural-net training on FPGA — .t27 spec → Verilog → silicon, bit-exact

📄 Full CV / resume: CV.md

Send your RTL — get it measured on a live Xilinx Artix-7, not simulated.

You receive a signed report: bit-exact conformance against an independent reference model (KAT vectors, not a testbench written from the same source), achieved timing and slack, resource usage, a latch-free check, and the bitstream — produced on a fully open-source flow (Yosys · nextpnr-xilinx · prjxray · openFPGALoader · iverilog), so you can re-run every number yourself without a vendor licence.

• 🌐 t27.ai/verification — full details, scope and terms
• 📄 Read a sample report — real GF16 4×4 matmul on XC7A200T
• 📊 t27.ai/proof — every measured result, and how it was verified
• 🔧 t27.ai/ip — license a core that has already been to silicon
• 🎓 t27.ai/course — the same method taught: train a neural network on an FPGA
• 💵 From $300 per core · 3–5 working days · first module free
• 🔒 Your sources are never published or reused, and are deleted on request. NDA welcome.
• 📬 admin@t27.ai

Useful if you publish open-source IP, are preparing a tape-out, or need "measured on hardware" numbers for a paper instead of simulation-only results.

I specialize in ternary ML — I design the number formats, the hardware that runs them, and the agent systems on top, from RTL to silicon.

• ⚡ FPGA RTL — custom float arithmetic, matmul cores, open-source toolchain (Yosys / nextpnr / XVC)
• 🧠 ML Infrastructure — CPU-optimized training, quantization, ternary neural nets in Rust
• 🤖 AI Agent Systems — multi-agent orchestration, MCP, Claude, RAG pipelines
• 📱 React Native / TypeScript — production apps and courses since 2016

• GoldenFloat: A Phi-Derived Static-Split Floating-Point Family from GF4 to GF1024 with a Lucas-Exact Integer Identity — arXiv:2606.05017
• An 83-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and Microscaling Formats — arXiv:2606.09686

• Google DeepMind AGI Hackathon 2026 — agi-hackathon evaluation framework
• The Open League (TON) — NeuroCalls meeting-analysis service
• OpenAI Parameter Golf — competition entry with novel GF16 quantization (Trinity Cognitive Stack): ~50% weight compression · ~3e-5 roundtrip error — parameter-golf-trinity
• DARPA CLARA (PA-25-07-02) — submission: Trinity Cognitive Stack; 10 CLARA reasoning gaps realized in open-RTL silicon (TinyTapeout SKY130, 3 chips) — trinity-clara
• 186 repositories — the Trinity ecosystem: ternary compute, FPGA, formal Coq/Rocq proofs, MCP servers, on-chain contracts

FPGA / Hardware    Verilog · Yosys · nextpnr · prjxray · openXC7 · XVC/JTAG
                   openFPGALoader · Vivado · iverilog · SKY130 PDK · TinyTapeout

Systems / Research Rust · Zig · Coq · LaTeX · GF16 custom float · ternary logic
                   FPGA-validated arithmetic · open silicon flow

AI / ML            Claude Code · MCP · RAG · multi-agent systems · LLM training
                   CPU-optimized inference · neural net quantization

Web / Mobile       TypeScript · React Native · React · Node.js · GraphQL
                   Supabase · Cloudflare Workers · Gleam

Blockchain / Web3  Solana · Ethereum · ERC-20 · DeFi · DAO tokenomics

_{φ² + φ⁻² = 3 · Building hardware that thinks}