T27.AI

About

From an arXiv paper to a fabricated chip.

I design number formats and the silicon that runs them. GF-T started as a paper, became RTL that maps into Artix-7 fabric with no hard multipliers at all, and went through a SKY130 tape-out — on a toolchain anyone can install for free. Before hardware I spent a decade building products and teaching: over a thousand developers, and the first React Native course in the Russian-speaking internet.

What I do

Custom arithmetic

Number formats designed against your constraints — ternary, low-precision, φ-based — each with an independent reference model and bit-exact vectors, not just a claim.

RTL to silicon

Synthesisable Verilog through an open flow: Yosys, nextpnr-xilinx, prjxray, iverilog. Measured on three Artix-7 boards I own, and taken to SKY130 when it needs to be.

Verification

Bit-exact conformance against models written from the specification rather than from the design — the only kind of check that can disagree with the RTL.

Teaching

Over a thousand developers taught. Hardware is the current subject; the method has not changed.

Published

arXiv:2606.05017 — GoldenFloat

A φ-based floating-point family, GF4 through GF1024, with the reference implementations that make it checkable.

arXiv:2606.09686 — 83 numeric formats

A conformance catalogue with bit-exact vectors for FP8, BF16, MXFP4 and microscaling formats, published so anyone can validate their own arithmetic against it.

Working with me

Remote, UTC+7

Based in Thailand, working with teams across Europe and North America.

Open tools by default

Nothing I deliver requires you to buy a licence to reproduce it.

Available

Contract and part-time hardware-AI, FPGA/RTL and ML-systems work.

The CV, the papers and the source are all one click away. If the work looks relevant, write — I answer.