Software engineer, data scientist & AI researcher
A deterministic deep-learning engine in Rust, written from scratch: tensors, reverse-mode autodiff and transformer building blocks (grouped-query attention, RoPE, a KV cache, mixture of experts, AdamW) on CPU, Apple Metal and NVIDIA CUDA. Runs are bit-for-bit reproducible on the same machine, and many seeds train in one pass. The numerical engine of Noetic.
The self-hosted agent OS — I'm the creator and maintainer. Run a fleet of AI agents on your own hardware, against any model, from one dashboard; peer-to-peer networked, nothing phones home.
Research on identifiable computational units (caps) for AI. Current paper: cap-augmented transformers achieve a 51% perplexity reduction over a matched-size baseline on TinyStories. Long-term direction: identifiable, self-evolving architectures.
The Limnx kernel - A small, hackable Unix-like OS for x86_64 and ARM64.