// ABOUT
Not a promise from a machine. Anomly builds the arithmetic, the runtime and the receipt format that let anyone prove what an AI computation actually did, on any hardware, after the fact.
b-posit operands with a 256-bit exact accumulator: a dot product rounds once, at the end, instead of once per term. Same weights, same inputs, same bits, on any hardware.
Every inference and every training step can carry a re-executable receipt. An independent implementation replays it from the weights and the prompt and compares bits. INVAR ships this today, open source.
The same datapath runs bit-exact on a Xilinx Alveo FPGA, on Tenstorrent accelerator kernels and on NVIDIA tensor cores at parity, and has been taped out on 130 nm, with the first die on our bench.
A small team that has shipped the whole stack, from the number format to a fabricated die.
Took posit arithmetic all the way to working hardware: the software implementation, the RTL and its formal proofs, the SKY130 tapeout, the silicon bring-up, the accelerator kernel stack, and the SDK. Previously spent a decade building GPU-compute infrastructure as founder of RevUp Render, the first real-time cloud-rendering platform. Inventor on multiple U.S. patents.
LinkedIn
HPC pioneer: creator of Gustafson’s Law, inaugural Gordon Bell Prize winner (1988), builder of the first commercial computer cluster, and inventor of the posit and quire arithmetic Anomly builds on. Independently reviewed and confirmed Anomly’s implementation as standard-compliant.
LinkedInAnomly, Inc. founded in the Washington DC area.
RaySpace: the first posit-based ray-tracing engine, built as a proof of concept that the arithmetic holds up in a real rendering pipeline.
b-posit arithmetic and the exact 256-bit quire implemented in software, RTL and CUDA; datapath blocks machine-proven correct for all inputs.
Exact GEMM bit-exact on a Xilinx Alveo U200; first SKY130 die delivered and brought up on the bench; implementation reviewed by John Gustafson, all 65,534 posit16 codes verified.
INVAR released open source: receipted local inference on the open Computation Receipts spec. Paper submitted to CoNGA 2026.
Whole-graph deterministic inference: identical bits on x86, NVIDIA CUDA and ARM, reproduced by two independent reference implementations.
INVAR and the Computation Receipts specification are Apache-2.0 on GitHub. Receipted models and exact-format GGUFs are published on Hugging Face.
For investor relations, use the contact form with "Investor" selected.
Media inquiries and introductions via the contact form. We hire people who have shipped hard things where AI systems meet silicon.
INVAR installs in one line on hardware you already own. Every answer comes with a receipt you can re-execute.
Try INVAR