Jeffrey V. Johnson, Ph.D.

Building infrastructure for model evaluation and agent systems

Resume actions Markdown AI/ATS

Summary

Ph.D. mathematician and systems builder focused on the infrastructure that makes agent-assisted development reliable: typed contracts, explicit permissions, reproducible environments, automated checks, and reviewable evidence. I apply these controls while building applications, data pipelines, evaluation systems, and infrastructure. My strongest fit is research engineering in agent systems, developer infrastructure, model evaluation, and safeguards.

Agent-development infrastructure

Host Capability Substrate : typed capability and evidence contracts

TypeScript · Zod · JSON Schema · GitHub Actions

  • Built 46 Zod entity schemas compiled to 67 generated JSON Schemas, with about 500 tests.
  • Modeled host capabilities, operations, execution contexts, policy decisions, provenance-typed evidence, approvals, leases, audit chains, and scoped authorization grants.
  • Wrote a 19-invariant governance charter and four-ring architecture, enforced by import boundaries and one CI gate for policy, boundary, secret, and schema-drift checks.
  • The public repository demonstrates the typed model and source boundaries, not universal host-level runtime enforcement.

Agentic Architecture Audit : specification and drift checks for agent-operated codebases

Specification · JSON Schema · Python tooling

  • Authored a two-stage audit-spec of 2,075 lines that keeps evidence collection separate from judgment.
  • Defined an 11-dimension rubric, 12 JSON output schemas, and a prioritization model based on severity, confidence, and weight.
  • Built a stdlib-only Python drift linter with 10 deterministic checks and a 27-case negative self-test in pre-commit and CI.

Governance and agent-control architecture : federated authority for agent-assisted work

An anonymized private case study · OpenTofu · OPA/Rego · TypeScript · Cloudflare · Hetzner · Proxmox

  • Designed a multi-repository architecture that separates human decision rights, agent permissions, identity and authorization, infrastructure enforcement, runtime admission, and repository-local authority.
  • Modeled agent authority as the intersection of mandate, available capability, and credential reach, with structured, secret-free handbacks for human judgment or privileged work.
  • Implemented selected controls including 26 OPA/conftest plan rules, repository governance as code, Zero Trust infrastructure, and scheduled drift checks.
  • Selected controls are operational; runtime admission is source-tested rather than deployed across the organization. This is a single-operator case study, not an enterprise-adoption claim.

Selected applications and evaluations

Agentic-Coding Evaluation Lab : fail-closed evaluation evidence

Public methodology and verification prototype · Python · Quarto · JSON contracts

  • Built and published a model-evaluation methodology and verification prototype that preserves evidence and withholds a recommendation when scientific prerequisites are incomplete.
  • Built a draft evidence contract and a separate verifier that recalculates a synthetic NOT_EVALUABLE outcome and rejects missing, inconsistent, or altered evidence.
  • Current evidence covers framework behavior and synthetic conformance, not a powered real-model result or promotion recommendation.

Budget Triage : financial evidence workbench

Public employer-facing source snapshot · active development private · Bun · TypeScript · Hono · PostgreSQL · not deployed

  • Built a financial evidence workbench for statements, receipts, and connected-account data with deterministic-first processing and human review.
  • Implemented exact monetary arithmetic, tenant-scoped atomic persistence, append-only decision records, authentication, Plaid, Infisical, and OpenTelemetry.
  • Built a synthetic known-answer and service-observation harness. Its current result is 6 passing and 9 pending assertions, so it remains incomplete and uncertified.
  • Published a sanitized, MIT-licensed source snapshot for employer review on 2026-07-21. Active development remains private; publication is not deployment or an open contribution surface. Explore the public source.

ScopeCam : native Android UVC camera application

Private Android · Kotlin · Jetpack Compose · C++20 · JNI/NDK · USB/UVC

  • Built a multi-module Android microscope application spanning Compose UI, Kotlin orchestration, a native C++ camera engine, USB lifecycle, persistence, diagnostics, and physical-device verification.
  • Diagnosed a camera-replug ANR across the JVM/native teardown boundary and implemented a device-verified recovery with the remaining leak tradeoff documented.
  • Delivered a signed client alpha; the private product is not broadly released.

Dicee : representative delivered software

Rust · WebAssembly · SvelteKit · Cloudflare Durable Objects · WebSockets

  • Built and deployed a family multiplayer game with a Rust-to-WebAssembly rules engine, real-time state through Cloudflare Durable Objects and WebSockets, and a SvelteKit client.
  • Public source and a live deployment demonstrate end-to-end delivery; traffic, performance, bundle-size, and current test-count claims remain withheld.

Technical skills

  • Languages: Python, TypeScript, SQL, Bash, Kotlin, C/C++, Ruby.
  • Contracts and controls: Zod, JSON Schema, OPA/Rego, OpenTofu, typed authorization, sandboxing, egress policy, provenance, and CI-enforced conformance checks.
  • Agent and evaluation systems: model and harness evaluation, evidence contracts, OpenAI-compatible gateways, provider routing, coding-agent workflows, and trace/provenance pipelines.
  • Infrastructure and data: PostgreSQL, Supabase, TimescaleDB, Redis, pgvector, Cloudflare Zero Trust, Hetzner, Proxmox, Docker, GitHub Actions, self-hosted runners, OpenTelemetry, Infisical, and 1Password.
  • Native and application systems: Android, Jetpack Compose, USB/UVC, JNI/NDK, CMake, and WebAssembly.
  • Mathematics and statistics: functional analysis, applied statistics, Wilson intervals, bootstrap confidence intervals, and paired sign tests.

Professional experience

Associate Professor of Mathematics (tenured)

Kenai Peninsula College, University of Alaska Anchorage · Homer, Alaska · August 2017-present

Previously American University in Cairo, 2014-2017; University of Montana, 2008-2014.

  • Teach undergraduate mathematics and statistics, including calculus and applied statistics; full teaching-and-service appointment; awarded tenure in 2023.
  • Consult for local non-profit organizations on statistical interpretation and study design.
  • Communicate technical and mathematical ideas to audiences with widely varying backgrounds.

Independent systems engineer and researcher (self-directed)

Self-directed; Happy Patterns LLC product entity · Homer, Alaska · 2023-present

  • Build and maintain a multi-repository engineering program for reliable, specification-driven agent development across applications, data pipelines, evaluation systems, and infrastructure.
  • Develop repository contracts, managed human and agent shell environments, typed authority models, and automated verification gates.
  • Designed and operate selected controls from an anonymized, single-operator governance and agent-control architecture spanning policy-as-code, IaC, Zero Trust networking, local and remote servers, and credential custody.

Education

Ph.D., Mathematics

University of Montana, Missoula, MT · 2013

Dissertation: Peripherally-Multiplicative Spectral Preservers Between Function Algebras.

Publications

Cadieux, J. M., Pyhala, S. L., & Johnson, J. V. (2023). Pediatric Walking Speed Normal Reference Values in a Local Population. Pediatric Physical Therapy, 35(3), 314-320. Led the statistical analysis.

Johnson, J. V., & Tonev, T. (2012). Spectral Conditions for Composition Operators on Algebras of Functions. Communications in Mathematics and Applications, 3(1).