thirdshift
An open-source Rust orchestrator that takes GitHub issues through implementation, review, and CI to pull requests, unattended.
My current focus is thirdshift, an open-source software factory in Rust, and the evaluations that guide its design. As Head of Engineering at Educational Travel Adventures, I also lead a two-person engineering team running reservations, payments, a Linux fleet, and AI-agent infrastructure for a multi-million-dollar travel business.
Software factories, the agent infrastructure around them, and the evidence behind their design.
An open-source Rust orchestrator that takes GitHub issues through implementation, review, and CI to pull requests, unattended.
183 replayed code-review sessions investigating where an additional reviewer helps an unattended software factory.
An open-source Rust launcher (va) that resolves vault secrets into a coding agent's own process, so Claude Code, Codex, Grok, Kimi Code, Antigravity, and Muse Code get real credentials without a pile of .env files on disk.
va run gives one-shot tools the same path. 32 releases through v0.4.25, for macOS and Linux.
Production systems with operational and revenue impact.
The core operations platform at Educational Travel Adventures - a full-lifecycle group travel ERP handling sales pipeline and lead qualification, itinerary building, reservations, vendor management, passenger logistics, payment processing, communications, and 70+ reports across finance, sales, marketing, and operations.
Tourbot Chat runs inside three Manager Sandboxes - role-specific Tourbot instances in Docker containers for Sales, General Management, and Marketing - giving managers natural-language access to business intelligence and a safe place to prototype platform changes.
Persistent Claude Code and Codex CLI sessions running inside each role's Manager Sandbox - hosted by the platform, reachable two ways: a browser terminal built into the ERP itself, and Claude's remote-control UI for driving the same sessions conversationally. Where Tourbot Chat gives managers an assistant, this hands them the same coding agents an engineer uses - with the platform supplying the guardrails.
An agentic engineering platform for Educational Travel Adventures: small, self-describing staff requests in; Factory Workers are agent attempts running in short-lived Firecracker microVMs; evidenced draft pull requests out. One human gate - the PR - held by engineering and the company owner. The production deploy path is never touched.
factoryctl) on a dedicated host with socket-only PostgreSQL, an append-only ledger, and a draft-only publisher under a GitHub App machine identity. Each agent attempt becomes a Factory Worker executing implement and verify work inside a short-lived Firecracker microVM, separate from the Docker-based Manager Sandboxes. Evidence capture, adversarial cross-family review, and a constitution (AGENTS.md) SHA-pinned in CI keep every run attributable. Phase 1 sessions covering isolation, verifier, publisher, dashboard, and the evidence lifecycle are merged on live infrastructure.
A mobile-first PWA for tour guide contractors at Educational Travel Adventures - itinerary and medical info access, document workflows, daily and summary reporting with live auto-save, expense tracking with receipt uploads, SMS messaging to travelers, calendar scheduling, and manager-side administration.
A Python/Flask operations console bringing fleet health, deployment review, branch previews, browser terminals, and agent credential management into one authenticated workspace.
The customer-facing side of Educational Travel Adventures - group leaders, parents, and travelers self-serve account creation, trip registration, payment plans, and communication. Inherited from a previous IT team and substantially rebuilt by me and the team I lead.
The cron-driven pipeline that turns scheduled customer payments into NACHA-formatted batch files and submits them to the bank for ACH processing - the system that actually moves the money. Inherited the original implementation; took ownership, modernized the runtime, and kept it running through platform migrations.
A six-round, profile-driven performance pass on Tourbot's group manifest page - the document staff print before every trip - taking page loads from 5-7 seconds to about 1 second by eliminating N+1 queries, batching in-request prefetches, and adding a targeted history-table index.
Booking, Price_statement, and ReturnPaySked constructors each ran helper queries per row of a 50-passenger group - almost all of it avoidable repeat work.
Booking::__construct skippable so 50 passengers hydrate from a single batched IN(…) query, and added a Price_statement::prefetch() layer - static caches and batched queries warmed once per request, with per-instance fallbacks for other callers. Replaced a leading-wildcard LIKE in the cancellation-history query with an Action_Type column and composite index, cutting EXPLAIN's rows-examined from 4,165 to 1.
An architectural rework of Tourbot's internal alerts page - a screen staff hit constantly throughout the day - that moves alert computation off the request path entirely. A background PHP daemon pre-computes every alert into a dedicated cache table, so the page just reads pre-shaped rows - cutting load times from 5-10 seconds to under 1 second.
A Prometheus, Grafana, and Alertmanager stack on a dedicated Terraform-provisioned droplet, instrumenting Educational Travel Adventures' 14-host fleet - followed by a measurement-driven tuning pass that removed ~80% of the production database's query load.
performance_schema-driven pass ranked 27 days of production queries by total time: three full-table-scan hot queries dominated, and three composite indexes - applied with online DDL (INPLACE, LOCK=NONE), EXPLAIN-verified before and after, rollback written first - eliminated them. The remaining load was diagnosed as application-level N+1 call volume and deliberately left for an app-layer fix rather than papered over with infrastructure.
A multi-year, codebase-wide modernization of Educational Travel Adventures' platform - PHP 5.6 to PHP 8, MySQL 5.7 to MySQL 8.4 LTS, and CentOS 7 to Rocky Linux 9 - done in production, without freezing the company.
mysql_* calls, declared properties on long-lived models, tightened type signatures. Shipped the MySQL 5.7 → 8.4 migration plan and the CentOS 7 → Rocky Linux 9 cutover, with deployment notes, schema-compatibility audits, and per-server runbooks. Coordinated cross-database export/import for customer accounts, dual-running old and new database servers during transition. Mentored the developer working alongside me on the same codebase.
Focused builds for payments, publishing, and straightforward business visibility.
A subscription payment system with Stripe integration for recurring revenue and secure transaction handling.
A marketing site for a catering business - clear presentation, service visibility, and lead generation.
A marketing and contact site for a faith-centered coach-mentoring practice - presenting one-on-one mentoring, seminars, and consulting for coaches at Christian athletic programs.
A curated set of personal builds, selected for relevance to platform and AI agent infrastructure work. The full shelf of published apps lives at stephens.page/apps.
A TypeScript telephone assistant that checks Google Calendar availability, reads appointment details back, and books a call only after explicit caller confirmation.
A local-first challenge tracker I built for my sister, who used it through her whole 75-day challenge. It brought her Google Sheet, separate water app, and workout screenshots into one place.
A Flask blueprint for managing long-running AI agents from a browser: it lists every tmux session across one host or a fleet, attaches live through xterm.js, and creates or kills sessions over SSH. It is built default-deny, with create-specific scopes, attach caps, and a documented threat model for handing out remote-shell access over the web.
A pull request to the php-mysql-replication library so its binlog client can complete the handshake against MySQL 8.4, whose default caching_sha2_password plugin left the existing code two packets out of sync.
Infrastructure-as-code for a personal web fleet of ~70 hostnames across 10 domains: DNS consolidated from four registrars onto Cloudflare and held as ~220 Terraform-managed records, with AWS S3 remote state, Ansible provisioning, and a plan-only disaster-recovery blueprint.
terraform plan reported zero changes, so no traffic-serving record was ever dropped or duplicated. Kept the state in AWS S3, deliberately off the compute provider, so a DigitalOcean outage cannot destroy both the box and the state that rebuilds it (S3-native locking, no DynamoDB). Wrote Ansible roles for one-command subdomain provisioning (Apache vhost plus Let's Encrypt, ordered so DNS resolves before the certbot challenge) and for rebuilding the base stack on a fresh host. Added a separate, plan-only Terraform module describing the droplet, volume, and firewall for disaster recovery, plus GitHub Actions running fmt, validate, and plan-on-PR.
The Kubernetes Demo was a separate single-node k3s learning and portfolio environment, not ETA production. It ran a live, HTTPS-secured two-tier app - a stateless web Deployment in front of a Redis StatefulSet - provisioned end to end by Terraform and cloud-init, until I tore the AWS node down on September 20, 2026.
FROM scratch image) fronting a Redis StatefulSet with a PersistentVolumeClaim, with production-grade manifests: a rolling-update Deployment with liveness/readiness probes, CPU/memory limits, a hardened securityContext (non-root, read-only root filesystem, dropped capabilities), ConfigMap and Secret injected via envFrom, a HorizontalPodAutoscaler, a Traefik ingress, and kustomize. Terraform plus cloud-init provision an AWS EC2 node and bootstrap k3s and the app; cert-manager and Let's Encrypt issue TLS. The manifests are schema-validated with kubeconform in CI, and a five-rule OPA Gatekeeper admission layer (non-root, resource limits, no :latest, the hardening triad, required probes) enforces the same posture at the API server - rejected-if-violated, not trusted-by-convention - with one rule re-expressed as a built-in ValidatingAdmissionPolicy in CEL to make the engine choice legible.
/count incremented a counter held in Redis and shared across both app pods - the whole path from terraform apply to a running, HTTPS-terminated cluster, as code. The screenshot is from a live load test; the application, manifests, and policy set stay public in the repository.
A couch word game for two to six phones and a laptop. Two far-apart words go up, everyone writes the word that bridges them, and a language model grades how close the table landed. The model writes the puzzle and keeps score; it never takes a player's seat.
game.ex with no timers, sockets, or model calls, so 1,102 tests grade against a fake Steward and never spend a cent. Every model output passes a deterministic gate before a player sees it: a schema validator, Elixir stemming that stops a Bridge from reusing a Bookend's root, and a prose linter on the Scribe's recaps. Phoenix LiveView gives each phone a private view, and the Chronicle is append-only Postgres.
A Flask web app that turns a Markdown document into an MP3 you can listen to, narrated by a neural text-to-speech model running entirely on my own server - no third-party API, and no text or audio ever leaves the box.
.md), pick a voice, and convert to a downloadable MP3. Synthesis runs on a self-hosted Kokoro-82M (ONNX) model in its own loopback-bound service, so no text or audio leaves the server and there is no external AI dependency or per-use cost. A Library keeps each rendering alongside its source-Markdown sidecar; sharing is via revocable capability-URL tokens (/share/<token>) - the only thing ever made public is a link you create and can revoke. Built as a seven-boundary safety demo: it never emails, posts, or acts on your behalf.
A privacy-first fertility-cycle charting app for the Creighton Model - local-first and offline, with auto-computed CrMS stamps, end-to-end encrypted sync, and provider sharing.
A waterfall white-noise player for focus, prayer, and sleep - one headless Rust core driving six native shells. Web (PWA), Android, Windows, and macOS are released; iOS and its watchOS companion are in TestFlight beta.
dispatch(command) → {snapshot, effects} boundary serialized as one JSON wire shape, so every shell is a thin native UI over shared Rust logic. Three bindings carry it: wasm-bindgen for the web PWA, UniFFI for Kotlin on Android and Swift on macOS and iOS, and a hand-rolled C ABI for C# on Windows. The watchOS app is a thin remote that talks to the iPhone. CI compiles every shell, and the macOS build is Developer ID signed and notarized by Apple.
An MCP server that gives AI assistants access to the Magisterium API, returning Church teaching with source citations.