Jeffrey V. Johnson, Ph.D.
Building infrastructure for model evaluation and agent systems
I build software and development infrastructure for agent-assisted work: clear specifications, explicit permissions, reproducible environments, automated checks, and evidence that can be reviewed.
My current work includes model-evaluation methods, agent controls, data pipelines, applications, and developer tooling.
Tenured associate professor of mathematics, Kenai Peninsula College/UAA · Homer, Alaska
Selected work
A few projects show the work directly. Each project page separates what is implemented from what remains open.
- Host Capability Substrate Public TypeScript implementation of explicit host capabilities, authority, policy decisions, and evidence. It includes 46 Zod entity schemas, 67 generated JSON Schemas, and about 500 tests. Examine the public implementation →
- Budget Triage Financial evidence workbench with exact monetary handling, provenance, tenant isolation, and human review. A public source snapshot is available; active development is private; the system is not deployed. Review the workbench →
-
Agentic-Coding Evaluation Lab
Public fail-closed model-evaluation method. A separate verifier recomputes a synthetic
NOT_EVALUABLEresult when required evidence is missing. It does not establish real-model performance. Explore the method and evidence → - Dicee Live multiplayer browser game using Rust/WebAssembly, SvelteKit, Cloudflare Durable Objects, and WebSockets, with public source and a live deployment. Review the delivered system →
More relevant work: Governance and agent-control architecture · Agentic Architecture Audit · ScopeCam · Full project index