A studio for agentic software.

Working together

Put agents to work, safely.

Most teams can see agentic development coming and aren't sure how to adopt it without shipping something they'll regret. I embed with your engineers to fix that: wiring agents into the real workflow, setting the guardrails, and building the security habits that keep an agentic codebase trustworthy.

The projects and research below are where I prove this out. It's the same work I'd bring to your team.

• Agentic development, in your stack and your pipeline.
• Security review for AI-written and agent-executed code.
• Hands-on with your engineers, not a slide deck.

What I'm building

Projects

• Jess — the strap on the falcon's leg

  Streaming agent harness with memory and skills.

  work in progress

  github →

• lmkit

  Train language models from scratch on one machine: pretrain, anneal, SFT, tokenizer, eval, and experiment tracking. Architecture-agnostic, and it runs the same on CUDA or AMD ROCm.

  work in progress

  github →

• lmkit-go

  From-scratch LLM training in pure Go on an XLA backend, with fused flash attention on both vendors. On NVIDIA it lowers to cuDNN fused SDPA, generalized into a fused-attention abstraction and upstreamed to gomlx/go-xla. On AMD it binds AOTriton flash kernels through an XLA custom-call and a PJRT FFI handler, so all models trains at 2k context on local consumer GPUs.

  work in progress

  github →

AI research

Research

Experiments and writeups, mostly on local inference and the internals of attention. Formal profile on Hugging Face →

• Differential Transformer on Apple Silicon

  Paired-init reproduction of differential attention in MLX with custom Metal kernels. An honest negative at small scale: no generalization edge over vanilla, cross-checked against the PyTorch reference.

  writeup → Hugging Face →

• Guardian: a dialect-fair hate-speech classifier

  Hate-speech classifier fine-tuned from ModernBERT-large that closes the dialect-fairness gap. Off-the-shelf models flag benign African-American English at over 2x the rate of General-American English; Guardian cuts that bias roughly 5x while holding 93% recall. The fix turned out to be volume of high-AAE-safe training text, not loss reweighting or adversarial methods.

  writeup →

More as I finish them.

About

Guy Grigsby

Software engineer, based in Denver. I build agentic systems and AI tooling, mostly in Go, and I help teams fold that same practice, and its security, into how they ship. The projects and research here are what I'm shaping now, and the proof of what I'd bring to your work.