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 working across native applications, data pipelines, web systems, and agent infrastructure. Builds verification tools, investigates failures across system boundaries, and directs and reviews agent-assisted implementation. Combines that engineering practice with published mathematical and statistical research, university teaching, and large-course coordination. Seeking research engineering work with meaningful ownership on a specialized team.

Agent-development infrastructure

Independent, unpaid projects developed alongside a full-time faculty appointment, using coding agents in implementation and test development.

Host Capability Substrate : typed capability and evidence contracts

TypeScript · Zod · JSON Schema · GitHub Actions · Public source

  • Built typed capability and evidence contracts with generated JSON Schemas and automated checks.
  • Modeled host capabilities, operations, execution contexts, policy decisions, provenance-typed evidence, approvals, leases, audit chains, and scoped authorization grants.
  • Defined architecture boundaries and a governance charter, with CI checks for policy, imports, secrets, and schema drift.
  • 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 · Public source

  • Authored a two-stage audit-spec that keeps evidence collection separate from judgment.
  • Defined a structured review rubric, JSON output contracts, and a prioritization model based on severity, confidence, and weight.
  • Built a Python drift linter that binds derived documentation to specification content, with negative self-tests 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 OPA/conftest plan rules, repository governance as code, Zero Trust infrastructure, and scheduled drift checks. The plan gate is workflow-level enforcement; it is not proven as a required status check across every repository.
  • The July 15, 2026 review found selected controls operational; runtime admission was 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.
  • Implemented context-specific data relationships with explicit transitions, tenant checks, database row locks, and stale-update protection.
  • Diagnosed an authentication test that failed on database reuse, tracing the interaction between rollback and persistent audit records.
  • Built a synthetic known-answer and service-observation harness. The July 15, 2026 review recorded 6 passing and 9 pending assertions, an incomplete and uncertified result.
  • 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 in June 2026; the private product is not broadly released.
  • Corrected a CI dependency gap that could report success after an upstream failure skipped the actual checks. Source change reviewed in September 2026.

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.

Additional systems

  • Agent-runtime experiments: Built Python/C request-guard and process-launch experiments with faulty senders, receiver-side checks, and paired event-order tests. Source and recorded bounded experiments support failure analysis; overall runtime acceptance remains open.
  • Fall course publishing: Built a registry-driven pipeline with isolated public output, checksummed manifests, and executable checks for broken links and accidental source publication.

Upstream contribution

Inspect Scout : reliability correction under review

Submitted a Python correction for provider-error diagnostics lost between spawned workers and the parent process, with regression evidence for the proposed change. I responded to maintainer scope feedback with further validation on September 12, 2026. The PR remains under review, awaiting a reply.

Technical skills

  • Languages: Python, TypeScript, Rust, SQL, R, Bash, Kotlin, C/C++.
  • 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

Kenai Peninsula College, University of Alaska Anchorage

Associate Professor of Mathematics (tenured), 2023-present

Assistant Professor of Mathematics, 2017-2023 · Homer, Alaska

  • Teach mathematics and statistics, develop course materials, and communicate technical ideas to audiences with varied backgrounds.
  • Developed a five-part introductory R module for applied statistics and online instructional materials for computational methods.
  • Led statistical analysis for a peer-reviewed study of 1,593 children and adolescents, collaborating with physical therapy researchers.
  • Contribute to institutional service and mathematics outreach.

American University in Cairo

Assistant Professor of Mathematics, 2014-2017

Taught calculus, differential equations, and advanced undergraduate independent studies; contributed to curriculum development and external-review implementation.

University of Montana

Post-doctoral Lecturer, 2012-2014 · Course Coordinator, 2013 · Instructor, 2008-2012

Coordinated two statistics lectures serving 480 students, supervised lab instructors and graduate teaching assistants, and developed teaching materials. Conducted doctoral research in functional analysis and co-authored a peer-reviewed mathematics paper.

Education

Ph.D., Mathematical Sciences

University of Montana, Missoula, MT · 2013

Dissertation: Peripherally-Multiplicative Spectral Preservers Between Function Algebras.

B.A., Mathematics

Humboldt State University · 2004

Service

U.S. Peace Corps, Republic of South Africa, 2005-2007; MathCounts coaching; Economic Development Council, Homer, Alaska, 2020-2022.

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).