aeryx
We build agentic systems, and we help teams learn to build them too. In Go, from Denver.
Working together
Most teams can see agentic development coming and aren't sure how to adopt it without shipping something they'll regret. We 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 we prove this out. It's the same work we'd bring to your team.
What we're building
Streaming agent harness with memory and skills.
work in progress
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
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
A semantic edit protocol for Go. Agents query the typechecked workspace and submit compiler-checked mutations instead of editing text, so an edit that does not typecheck cannot reach disk. Invalid edits are unrepresentable, not merely flagged.
work in progress
AI research
Experiments and writeups, mostly on local inference and the internals of attention. Formal profile on Hugging Face →
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.
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.
More as we finish them.
About
A software studio based in Denver. We build agentic systems and AI tooling, mostly in Go, and we help teams fold that same practice, and its security, into how they ship. The projects and research here are what we're shaping now, and the proof of what we'd bring to your work.
Contact
Bringing agents into your team, hiring for a project, or just curious what we're up to?