AI engineering
From choosing a model to running it in production, on infrastructure you control.
What we build
- Model evaluation and selection. We test candidate models against your actual tasks and data, and measure quality, cost and speed before you commit.
- Fine-tuning. Data preparation, training runs, evaluation and handoff for a model trained on your data.
- Assistants trained on your workflows. An assistant that follows how your team actually works, with the prompts, tools and test suite to prove it does.
- Model deployment. Self-hosted models on Kubernetes, in your cloud account or on hardware we manage, including routing each request to the cheapest model that can handle it.
Recent work
- model-router: an LLM gateway that decides, per request, whether a local or a hosted model should answer, using a small trained decision model instead of asking an LLM.
- beads-worktree-orchestrator: runs several AI coding agents in parallel, each in its own isolated copy of the code, coordinated through a shared task queue.
- Local model serving on rented GPU hardware, and an accepted fix to ECA, an open-source AI coding assistant.
Projects are quoted as a fixed bid after an assessment, or as time and materials when the scope is still moving.
Not sure where to start? An AI Readiness assessment is the usual first step.