John Pritchard

I build systems that run AI on hardware you own.

What I've shipped

BIFROST / RAMEAU

A multi-model AI inference platform I designed, built, and operate solo.

  • Three nodes, twelve local routing tiers, a Kubernetes control plane.
  • A complexity classifier routes each request to the smallest model tier that can handle it.
  • LangGraph agents decompose tasks into parallel subtasks and call tools.
  • A tiered verification harness gates every output before it ships: emptiness and refusal checks, stub detection, acceptance-criteria coverage, semantic-embedding drift, and a blocking sandboxed unit-test gate, with an advisory compile check on the assembled result. Only the unit-test gate blocks; a failure there escalates with its evidence attached.

Over a measured 30-day window it moved 33.9 million tokens — a median of about 408,000 a day, peaking at 5.3 million on its busiest day. Local hardware carried 69% of that volume; most of the remainder ran on free-tier provider-hosted models, which held metered inference spend for the entire window to about six cents. It also supports a fully offline routing mode that drops every provider tier, for environments where nothing may leave the building. RAMEAU will be the containerized Apache-2.0 release of that platform, built for on-premises deployment.

Other production systems

  • A 27-endpoint document-generation service on an 8-container stack, running in production with zero cloud dependencies.
  • A self-hosted private cloud on 10 TB of software-defined storage behind a WireGuard mesh, with 99.9% uptime over three years.

How I work

I build with coding agents as my default mode, not as an experiment. Agents are dispatched in parallel, each on its own branch and session, against a codebase I deliberately made agent-ready:

  • A written ruleset derived from real failures
  • A dispatch journal so agent work stays auditable
  • Protected-file conventions
  • Verification gates that run before anything merges

The governance is the interesting part. Running agents is easy. Running them so you can trust what comes out is the actual work.

Background

  • Thirty years maintaining the embedded control systems inside pipe organs: relay logic, low-voltage switching, electromechanical integration.
  • A doctorate in music.
  • Years teaching at two universities.
  • Learned BASIC on a VIC-20 in 1984, and came back to building when local models got good enough that one person could ship real infrastructure again.

That path means I think about systems the way a researcher approaches an experiment: rigorous documentation, disciplined iteration, honest failure analysis. It also means I can explain a technical system to someone who doesn't share my vocabulary, which turns out to matter more than it should.

Reach me

Email
jhpritch@hotmail.com
Links