GOODYEAR, AZ · REMOTE (US)
HYBRID PHOENIX
Jordan Truong, AI Engineer in Goodyear, AZ. Open to remote US roles or hybrid Phoenix.
Agentic engineering, built to written specs and gated by CI and evals before it ships.
Production agentic AI: multi-agent systems, RAG pipelines, and the infrastructure beneath them. Solo-built and run in daily production, end to end, from a terminal in Arizona.
The arc: twelve years a salon owner-operator. Then roughly four years in enterprise IT. Now AI engineering, shipping software that runs itself. Same instinct throughout: make the hard thing dependable.
Also live and provable: /ops — Langfuse-traced cost dashboard with per-request pricing · Anthropic Academy ×16, Claude Partner — Claude Code, and Google AI Professional Certificate.
Solo-built multi-agent VPS, SoloInvoice SaaS, RCM knowledge assistant. All in production.
Built production AI tooling across two enterprise campaigns: HIPAA RCM copilot (every scenario audited, 72 KB-grounded scripts, 27 RPM playbooks, 41 canned-note codes) and ACES Aid (162 flows, floor-wide adoption, 1,500+ indexed keywords). Both in daily floor use.
Managed provisioning across 1,000+ franchise locations. Led network/security audits with firmware rollouts sequenced across timezone waves.
Two enterprise campaigns: HMH education software (app/account triage, escalation) and Dell Technologies (hardware diagnostics via Zendesk, live Teams collaboration with senior techs, warranty logistics).
Twelve years running a service business end-to-end: hiring, payroll, scheduling, the P&L.
Open-source MCP memory server. One command to install: pip install all41n14lla. A 90-test suite runs green in CI across Python 3.11 through 3.14.
Ten agentic systems in daily production: healthcare + telecom copilots (every scenario audited, 162 flows), five in-browser demos (RAG, SQL, email triage, agent orchestration, semantic layer), SaaS with offline licensing, a 7-agent outreach engine, and the platform wiring all of it. Tap to browse.
Fifty-plus skills loaded on demand, fifteen MCP servers, auto-memory across sessions, and layered guardrails. The model is 10% of the system. The harness is the other 90%.
Scout finds the lead, Diagnoser maps the gap, Builder ships the fix, Filmer records the walkthrough, Checker validates it, and Pitcher sends the note. No staff and no retainer, at about $0.10 a lead.
Type a question in plain English. Watch it become a safe, parameterized query.
Paste raw email. The agent classifies, prioritizes, and drafts a reply in one pass.
Inline citations and confidence-gated abstention: when it is not sure, it declines. Domain: healthcare revenue cycle.
A three-role pipeline over real retrieved facts. The Critic checks the draft against the facts and forces a revision round when it finds an unsupported claim, capped at two rounds so it always finishes.
Ask a business question. The model can only pick a metric off a fixed, published catalog and never writes its own SQL. Questions outside the catalog are refused.
Or open the chat bubble (bottom-right) to ask me anything.
Every demo ships through typecheck, eval, and prompt-regression gates on every commit. You can watch the guardrails work: RAG declines instead of guessing, the Critic rejects unsupported claims, the semantic layer refuses metrics that are not in its catalog.
Fine-tuned Qwen2.5-0.5B locally on 8 examples from this site's own content, testing whether it could drop the ~190-line system prompt the live chatbot runs on. It couldn't. Real transcript from the eval run:
Base model + system prompt (correct): "An open-source MCP memory server for AI agents, maintained by Jordan Truong, published on PyPI and GitHub."
Fine-tuned, no prompt (hallucinated): "...a web application for managing all 41 million users of the Netflix app..."
Owner-operator of a salon and beauty business. Scheduling, payroll, inventory, hiring, client experience, all of it. The job taught systems thinking before I had a name for it: every constraint is a design problem, every bottleneck is a process failure waiting to be solved.
Joined the workforce-management tech stack at scale. Claims processing, healthcare RCM, multi-team ops coordination. Learned what large-system reliability actually means when a misconfigured rule costs someone a reimbursement. Then LLMs became capable enough to matter.
Shipped a PyPI MCP server, a live RAG+GraphRAG knowledge assistant, a multi-agent VPS brain, and a licensed SaaS product, all of it solo and all of it in production. The plan is to keep shipping until the systems are earning while I sleep.
One diagram, drawn from the code on the day this was written. Two entry points, one shared trust boundary, and three honesty tiers on every demo answer.
What this system is NOT: no PHI, no real user accounts, no payment data. The HIPAA-safe tool built at Valor is a separate system and is not shown here.
Five decisions from the systems above, with the cost attached. Anyone can list what they shipped. These are the choices that had a price, including the ones that cost me something.
Remote-US or Phoenix metro only. Open to full-time AI / agentic-systems roles.