Projects

Agent Neo

Agent Neo is my open-source agent for investigating infrastructure incidents. It’s one Go binary that drives the tools already on your machine (gh, kubectl, aws, ssh, database clients) and can open a PR when it finds the fix. brew install open-auto-intel/tap/neo. The write-up.

TensorWire

A GPU toolkit for training and running models, written in Go. It started as a weekend project because I was sick of fighting Python to fine-tune things. The repos below are public. The newer work is private until it’s ready.

• ai

  One CLI to train, fine-tune, chat with, quantize, and serve models. brew install tensorwire/tap/ai

• mongoose

  The GPU engine underneath, with Metal, CUDA, Vulkan/WebGPU, and CPU backends.

• gguf

  Reads and writes GGUF, SafeTensors, and NumPy files in pure Go.

• tokenizer

  A BPE tokenizer that loads HuggingFace and SentencePiece formats.

• helix

  My first fine-tuning optimizer. It couples paired weights (Q with K, gate with up).

• tesseract

  A desktop front end for ai serve.

Papers

• Sparse-First Training: skip the work for weights that aren’t learning. It measured 1.54× faster than PyTorch DDP on 2×H100, but that was against default DDP settings, and I owe PyTorch a fairer fight.
• SQ4, retired: a 4-bit weight format. Standard formats beat it once I measured properly. What I got wrong. I’d also opened a llama.cpp PR for it, and closed it when I retired the format.

Other stuff

• BlackBoxDelta: my automated forex trading platform, now retired. Thousands of transformer models forecasting signals across 40+ currency pairs, and an LLM writing up the market analysis with nobody editing it.
• schwab-trader: a Go client for the Schwab developer API.
• Ghost VPN (ghost-vpn.me): a VPN I started during the 2020–2021 lockdowns. You set it up by scanning a QR code and paid with PayPal or Ethereum. It got four paying customers in six months, and taught me a lot about how hard selling things is.
• Older open source: Puppet modules, an SVN-to-GitHub migration tool, and a pile of infrastructure tools on github.com/swizzley.