Scientific software, local AI, and the infrastructure under both.
I take on contract work in three areas, drawn from twelve years of building and running systems at observatories and from running open-weight models on hardware I built.
Scientific software and data infrastructure
Instrument control, acquisition, and the pipelines that carry raw output through to publication-ready results. Much of the work is the systems around the algorithm: built to run unattended, around the clock.
| Typical scope | Real-time and batch pipelines, instrument interfaces, data products and their archives, reprocessing campaigns. |
|---|---|
| Handover | Runbooks written for your staff, so your team owns the system from day one. |
Local AI on hardware you own
Open-weight models on your own hardware, for groups whose data can't go to a cloud API, or who want predictable costs and reproducible results. I work hands-on with the leading open source inference engines, vLLM and llama.cpp: benchmarking them, patching them, and running them daily for coding agents on GPU servers I built.
| Typical scope | Sizing and building an inference server, choosing between vLLM and llama.cpp, evaluating models on your own tasks, benchmarking, and connecting coding agents and tools to it. |
|---|---|
| Coding agents | Spec-first workflows, with checks a machine can run and a person reviewing every merge. |
| Handover | A server your team runs, benchmarks for your own workload, and a written record of what was tried and why. |
On-prem infrastructure and disaster recovery
Compute, storage, and virtualization on an open source hypervisor in your own building, for teams whose data may not leave the premises, whose instruments can't depend on an outside network, or whose funding makes capital easier to spend than a subscription. And a recovery plan that has been executed rather than written: a backup nobody has restored is a rumor, so the deliverable is the restore, performed by your people.
| Typical scope | Rack and power planning, hypervisors and storage, network and remote access, capacity planning, monitoring and alerting, second-site design, and restore drills. |
|---|---|
| Handover | Your own staff operating it, with a restore drill they repeat on a schedule, and me reachable rather than required. |