Justin Higgins

Software engineer. Twenty years in regulated industries, the last several spent running a small fleet of coding agents and keeping every receipt. Kansas → Oregon, remote.

jchigg2000.dev@gmail.com · LinkedIn · GitHub

Strengths, as Claude sees them

Written from the essays on this site and his session files: 3,738 Claude Code sessions since April 2026 and the 12,263 helper agents they started, counted by Claude from the files on 2026-09-26.

  • He reads his own instruments for lies. Three times a measuring tool he built told him something flattering. Each time he traced it back to the tool instead of taking the win, and wrote it up.
  • He verifies the artifact, not the process. A green check does not count until he has looked at what the far side actually received. He wrote up the time a privacy filter he put in front of a small app took it down for days while its health check stayed green.
  • He kills his own claims. He handed all 24 findings in one first draft to a reviewer told to refute them. Nine were refuted and fifteen overstated. He cut them, kept what he could read straight off the logs, and published the tally.
  • He runs agents unattended and keeps their state in a file. When one of those runs died partway through, a fresh session read the file and finished the job.
  • He says what he is not before you ask. Go, TypeScript, and Python daily. Java in production history, not daily. Container images and services rather than Kubernetes fleets. The agent runtime he built serves one person.

What I've built

Sole-author, public unless marked private. Counts as of 2026-09-26.

  • air-traffic - enterprise AI control plane in Go, standard library only. Policy engine, spend plane, config push with drift readback across sixteen vendor surfaces, inline redaction gateway. 189 tests.
  • piidetect - PII/PHI detection and redaction for Go, extracted from air-traffic's redaction gateway: standard library only, regex matches confirmed by checks like Luhn and IBAN mod-97, and a self-hosted Presidio sidecar as an optional tier. 34 tests.
  • gitjob.io - the job-search engine from The Principal IC Moat, private. Go, Postgres, pgvector. 193 test functions. Its public page at gitjob.io is the attestation spec below, not the job tool.
  • ai-attestation-spec - an open spec for signed receipts of AI-assisted work. A Receipt Nobody Issues Yet says where it stands.
  • mullion - a notarized macOS window manager in Swift. 81 tests. The weekend in Git Could Not Tell Me Who Wrote It.
  • claude-code-dev - 42 commands and 18 skills for Claude Code, kept as versioned specs. MIT.
  • claude-ace - a CLI for token and cost telemetry across agent sessions.
  • claude-code-log - a local, read-only dashboard over Claude Code transcripts and the repos they ran in, with estimated spend, full-text search, and the tab from Words That Mattered. 149 tests.
  • The agent operating layer, private, tour on request: an always-on dispatcher that queues and launches sessions, a searchable archive of every session, and a scorer that grades the work on an 11-category rubric.
  • Meadowlark IQ - a deployed question-answering system over 1,008 near-identical insurance contracts that went from 1% to 98% end-to-end accuracy, with typed refusals when the evidence is not there. Everything Looks Like Everything is the write-up.
  • effigy - privacy-preserving code husking so a model can review code it is not allowed to see. 69 tests. The Shape Leaks is the negative result.

Interactive pages built from the same records are in the showcase.

Meadowlark, my portfolio

Meadowlark is my portfolio, not a company: sole-author healthcare systems on synthetic data, running on public URLs so each claim can be checked. The products sit behind an access-token door, and a tour is on request. The docs say how each one is built.

The work I like

  • Staff or Principal IC, hands on. The keyboard stopped being the bottleneck. I want to stay on it.
  • Agent platforms, developer productivity, governed AI. The runtime, the evals, the receipts. The part of the job where the question is what did it do, and can we prove it.
  • High-stakes domains. Places where the cost of being wrong is real and most of the work is judgment. I have spent a career in them and I am not looking to leave them, just to move faster inside them.
  • Remote, from Kansas, moving to central Oregon. Companies that move at the pace the tooling now allows.
  • Not a title that means six different things at six companies. The Principal IC Moat is the long version.

How I work

Three essays that show it better than a bullet can.

Reactions, questions, roles: jchigg2000.dev@gmail.com.