Field Notes

Field Notes

Technical reports from building and running research computing

Dated reports on real work: what was built, what it measured, what broke and what was learned. Each one names its repositories and specs so the claims can be checked.

  • 2 November 2026

    Sensitive Data, Shared Cloud

    A paper and companion book that check a plan for a shared Google Cloud environment for HIPAA and FERPA research data against peers, regulations, vendor documentation and code, before anything is built.

  • 29 October 2026

    Read-Mostly: AI Assistants and Real Slurm Jobs

    How ursa-bifrost gives AI assistants typed, audited access to a Slurm cluster, why 219 mutation guards missed checks written wrong, and the fork ClusterWand.

  • 29 October 2026

    An AI Gateway for Research Labs

    A pilot AI gateway for research labs: open-source LiteLLM on Cloud Run, lab allowances, per-key caps, claim-link keys and a kill switch, and what broke.

  • 29 October 2026

    A Campus Petabyte

    How UC Riverside built CephRDS, a 3.2 PB NSF-funded Ceph object store, split it between campus research and the NRP, and what is still unproven for researchers.

  • 29 October 2026

    Governing 176 Research Cloud Projects

    How UC Riverside Research Computing governs 176 Google Cloud projects with budgets shown as allocations, a throttle at the cap and a read-only audit tool.

  • 29 October 2026

    A Seneschal for the Research Computing Desk

    How one AI agent drafts my Research Computing email and tickets, which approval gates are enforced in code, which are only written rules, and what the log says.

  • 29 October 2026

    Nexus: A Research Computing Relationship Ledger

    How Nexus, a PostgreSQL knowledge graph and interaction ledger, grew over 10 months into the record UC Riverside Research Computing and its AI agents work from.

  • 29 October 2026

    OmNet: A Homestead Network Where Code Decides

    How a homestead network runs on compiled sentinels and a small computer, and asks a model only about new log lines, under caps, a cache and my overrule.

  • 27 October 2026

    Directing a Fleet of AI Builders

    A dozen apps started on five machines with four AI models over one long weekend. What finished, what broke, and the rules that came out of it: the agents were capable; the failures came from what they were allowed to touch.

  • 6 October 2026

    A Road to Nautilus in One Day

    nrp-mcp, an open-source MCP server for the NRP Nautilus Kubernetes cluster, went from a written plan to a public, cross-platform release in one day; then five adversarial reviews said what was not ready yet.

  • 3 October 2026

    One Day on the Research Stack

    38 releases in one day across a cloud Slurm cluster, its MCP gateway, an AI research engine, an operator desk and a notes server: cheaper partitions, an always-on check partition, an outage fix, a migration off SSH, workspace separation, and a cluster that installs its own software catalog.