Donald B. Havery

Applied AI Engineer, Forward-Deployed Focus

Donald B. Havery

APPLIED AI ENGINEER · FORWARD-DEPLOYED FOCUS

AI systems
engineered
for production.

I take vague AI asks to production: agent orchestration, RAG, eval gates, and security controls, handed off clean. Applied AI systems engineering is the layer around the models that makes them shippable.

You hire me and the team of AI agents I build and run. They come with me.

SELECTED WORK RÉSUMÉ RESEARCH PAPER GITHUB LINKEDIN CONTACT
Research paper Reference-relative believability Independent, archived on Zenodo · DOI: 10.5281/zenodo.21138188
Works across
and more
Shipped & live

FOUR BUILDS YOU CAN CHECK.

The strongest evidence is public and measurable: a gateway with failover, an eval gate that exits non-zero on regression, and an agent harness that records retry, fallback, and recovery paths. The fourth is not a repo. It runs a live business, answers the customers who arrive, and takes payment. Inspect the repos, reports, and tests, or go and talk to the agent.

angrynirds concierge live transcript: get_price for the ecommerce storefront tier returns from $5,500 one-time, get_timeline returns 6 to 8 weeks, the guard scans the reply for un-tooled currency figures and finds none, and the agent talks around a card whose digits it never types
LIVE · IN PRODUCTION

CUSTOMER-FACING AGENT / GUARDRAILS

Production AI agent running a live business

  • Tool-grounded pricing
  • Un-tooled figure guard
  • Grounded retrieval
  • Live Stripe checkout

The front desk of my own business, not a demo. Every money figure a customer sees comes from a price-book tool call, and a stream-level guard stops any un-tooled figure from completing in visible text, then hands the conversation to a human. Policy answers quote source excerpts or say they do not know. Checkout is live Stripe.

TALK TO IT → Live · angrynirds.com
citadel gateway failover trace: Router.resolve_all returns the chain anthropic to bedrock to ollama, three anthropic failures open its circuit breaker, the open circuit is skipped, and the request is served by bedrock at $0.0032 per 1k tokens
PUBLIC · CI GREEN

LLM GATEWAY / PLATFORM RELIABILITY

LLM reliability gateway

  • Provider routing
  • Cross-provider failover
  • Circuit breakers
  • From-scratch HNSW search

An OpenAI-compatible control layer for LLM traffic: provider priorities, circuit breakers, semantic cache, and from-scratch HNSW search. The proof is deliberately inspectable: public code, MIT license, green CI, and cloneable tests.

OPEN REPO → MIT · github.com/dbhavery/citadel
eval-gate HTML report: GATE PASSED, 100% pass rate against an 85% threshold, 13 of 13 cases, 48 of 48 checks, deterministic provider
PUBLIC · CI GREEN

RELEASE-QUALITY EVALS / RAG

Evaluation release gate for LLM & RAG

  • Prompt assertions
  • Retrieval precision / recall
  • Citation grounding
  • Non-zero exit on regression

A release gate for LLM/RAG behavior: prompt assertions, retrieval metrics, citation grounding, and baseline comparison. It runs offline and deterministic, then exits non-zero when a case regresses.

OPEN REPO → MIT · github.com/dbhavery/eval-gate
Agent orchestration reliability harness trace report: a run that plans, calls tools, retries a flaky tool with backoff classified as transient, then recovers to a successful result
PUBLIC · CI GREEN

AGENT ORCHESTRATION / RELIABILITY

Agent orchestration reliability harness

  • Typed tool schemas
  • Retry & fallback
  • Failure classification
  • Unsafe-action gate

A deterministic agent loop built around production failure modes: typed tools, structured traces, bounded retries, fallback, unsafe-action blocking, and explicit failure classes for timeout, malformed output, missing context, and unsafe requests.

OPEN REPO → MIT · github.com/dbhavery/agent-harness

Applied AI engineering, end to end.

The four builds above are the fast proof. The broader portfolio maps that proof to the capabilities senior AI teams hire for: architecture, orchestration, release quality, security posture, product surfaces, and infrastructure.

Agentic orchestration

Typed tool use, multi-step loops, retries and fallback, failure classification, and safety gates.

Proof: agent reliability harness, LLM reliability gateway, local-model MCP bridge

LLM platform architecture

API boundaries, model gateways, semantic cache, retrieval, ingestion, tracing, and cost control.

Proof: self-hosted LLM reliability gateway

Evaluation and reliability

Prompt assertions, RAG metrics, citation grounding, regression gates, and test reports.

Proof: LLM & RAG evaluation release gate, agent reliability harness

Security and governance

Control-to-evidence mapping, verdicts, audit logs, redaction, severity labels, and remediation.

Proof: policy-control evidence review, compliance gap-analysis app

Product, SaaS, and web design

Workflow design, dashboards, role-based portals, launch copy, and responsive UI.

Proof: fuel-delivery operations platform, live product sites

Infrastructure and bare metal

Docker, Linux, GPU workflows, local model bridges, IaaS choices, and systems that run without babysitting.

Proof: local-model MCP bridge, consumer-GPU fine-tuning pipeline, text-to-CAD loop

Languages. Python · JavaScript · TypeScript · HTML · CSS · SQL · Bash · PowerShell · Java · Kotlin · Rust · C++

AI engineering. Agent orchestration · Tool use and MCP · RAG and retrieval · Prompt and eval design · Regression gates · Failure classification · LoRA and QLoRA fine-tuning · Local inference · Speech to text and text to speech · Real-time talking-head avatars · Vector search

Models and platforms. Anthropic · OpenAI · Gemini · Hugging Face · Ollama · AWS · Google Cloud · Vercel

Product and web. React · Next.js · Node.js · FastAPI · Tailwind · Vite · Stripe · Cesium · Responsive and accessible UI · Design systems

Data and storage. PostgreSQL · SQLite · Redis · ChromaDB · Pydantic · Schema design · Migrations

Infrastructure and quality. Docker · Linux · GPU workflows · CI pipelines · Playwright · Vitest · pytest · Observability · Audit logging · Secrets handling

Foundations. Thirty ACE college-credit courses behind the applied work: Python, Java, web development, relational databases, networking, and statistics.

HOW I RUN A BUILD.

I am strongest where most AI prototypes fail: turning unclear scope into an architecture, product workflow, implementation plan, eval strategy, security posture, and clean ownership handoff.

ARCHITECTURE Provider boundaries, data flow, observability, and failure modes.
ORCHESTRATION Agents, tools, retrieval, memory, routing, and human handoffs.
QUALITY Evals, regression tests, release checks, and measurable behavior.
SECURITY Control evidence, audit trails, access boundaries, and review paths.

SIX MORE SHIPPED SYSTEMS.

Security and compliance review, operations tooling, live-data interfaces, local inference, and a live training site, each with public code or a live surface to inspect.

PUBLIC Self-correcting text-to-CAD loop Plain-English part requests become OpenSCAD, render, and get inspected by a vision model that feeds failures back into repair. Local execution, no API keys, and a compact deterministic test suite. OPEN REPO → Public · MIT · github.com/dbhavery/textcad PUBLIC · CI GREEN Policy-control evidence review Maps control statements to supplied evidence, returns per-control verdicts with matched excerpts, redacts secrets, and writes an append-only audit log for reviewer handoff. OPEN REPO → Public · redaction · reviewer decision aid LIVE DEMO Compliance gap-analysis app Runs representative-data gap analysis against SOC 2, GDPR, and HIPAA requirement sets, then generates policy and audit-report drafts while leaving the compliance decision with a human. TRY IT → Live Hugging Face Space LIVE DEMO Role-based fuel delivery operations platform Role-based dispatch, customer, driver, franchise, and HQ workflows for an on-site fuel delivery business, paired with a live public marketing site. Pick a role in the demo and walk the dashboards. TRY THE DEMO → StaFull · stafull.com is live · demo needs no account PUBLIC · DESKTOP APP Weather workstation on a live globe A physically real Earth carrying radar loops, satellite imagery, forecast models, station observations and severe-weather warnings, alongside planes, ships, satellites, quakes and fires. Local-first, single window, no cloud. OPEN REPO → Public repo · Cesium desktop app LIVE SITE Free 52-week CDL course with a published syllabus Trucking Tutor: the classroom half of a commercial licence, free, one lesson a day for a year. The full syllabus is live and every lesson names the regulation it comes from, drawn from the federal rules, the model CDL manual and all fifty state manuals. Written by a working driver and trainer, with the videos in production. OPEN THE SITE → Live · truckingtutor.com · 60+ pages

More systems across agents, voice, local inference, fine-tuning, and scheduled automation.

  • Local-first desktop AI assistant architectureReference: Aether
  • Push-to-talk dictation with on-device Whisper transcriptionReference: Vox
  • Multi-agent turn-taking on one shared microphone, arbitrated to a single speakerReference: Atrium
  • Grounded support-experience concept page, a self-playing transcript that cites its source and escalatesReference: support.dbhavery.ai, design demo
  • MCP bridge for local models, embeddings, vision, and memoryReference: McPlex
  • Local QLoRA fine-tuning pipeline for consumer GPUsReference: FineForge
  • AI pull-request review GitHub appReference: CodeRev
  • AI video and research triage pipelineReference: Frontier Scout
  • Scheduled AI news, system update, and social-posting automationReference: Morning Intel
  • Local image generation and inpainting desktop toolReference: Image Generator

EXPERIENCE.

One line of work stated three ways: build the thing, run the business that depends on it, and stay accountable for what it does after handoff.

  1. Applied AI Engineer, Forward-Deployed Focus Independent · Current

    Audit the workflow, prove it with evals, deploy it with an audit trail someone else can check. The work above is the record: public repositories, cloneable tests, green CI, and reports that exit non-zero when a case regresses.

    • Claude Code as the daily pair programmer, with prompts and evals in version control like any other production code
    • Reusable Claude Code skills and subagents for that workflow, version controlled
  2. Founder and owner-operator Built and run solo, alongside full-time work

    Four businesses carried end to end, from the operations software to the invoices: a role-based fuel-delivery operations platform with a live public site, a direct-to-consumer premium grooming brand, independent investigative case work, and angrynirds, a web and AI studio whose front desk is the production agent above, live and taking payments. The fifth venture is Trucking Tutor, a free 52-week CDL course: the full syllabus is live, every lesson names the regulation it comes from, and the videos are in production.

    • Dispatch, customer, driver, franchise and HQ workflows, designed and shipped
    • Brand, storefront, launch copy and fulfilment, owned personally
    • Trucking Tutor curriculum, site and channel, built from the regulations and all fifty state manuals
  3. Certified entry-level driver trainer DOT / FMCSA regulated freight

    Years in safety-critical regulated freight, where the documentation and the audit trail are the deliverable and a missing record is a violation. It is the reason evidence, not assertion, is how the engineering above is built.

CREDENTIALS ON RECORD.

Eighteen AI certifications and thirty ACE college-credit courses. The Coursera and Credly entries verify online at the links below; the Anthropic Academy set covers Claude, Claude Code, MCP, Bedrock, Vertex AI, and the AI Fluency track, and the IBM track is a ten-course professional certificate in AI product management.

Anthropic, the instructor track TEACHING THE AI FLUENCY FRAMEWORK Delivered by Anthropic with University College Cork, Ringling College of Art and Design, and Ireland’s Higher Education Authority and National Forum for the Enhancement of Teaching and Learning. April 2026. Anthropic’s course for the people who teach the framework, not only work in it. ABOUT THE PROGRAM → IBM AI PRODUCT MANAGER Professional certificate, ten courses, issued by IBM and SkillUp via Coursera. Product management lifecycle, AI product development stages, generative AI foundation models, and prompt engineering. Sep 2026. VERIFY ON COURSERA → 14× MORE FROM ANTHROPIC Claude 101, Claude Code in Action, Claude Code Skills, Building with Claude, MCP and MCP Advanced, Claude with Bedrock, Claude on Vertex AI, Claude Cowork, AI Capabilities & Limitations, and the AI Fluency track for educators, students and nonprofits. Mar to Apr 2026. ABOUT THE PROGRAM → GOOGLE AI Fundamentals, issued by Google via Coursera, with identity-confirmed verification on record. VERIFY ON COURSERA → DEEPLEARNING.AI Generative AI with Large Language Models, authorized by DeepLearning.AI and AWS, via Coursera. Mar 2026. VERIFY ON COURSERA → 30× ACE CREDIT Thirty ACE-recommended college-credit courses via Sophia Learning: Python, Java, web development, relational databases, networking, and statistics. BADGES ON CREDLY →

Education

Capella University B.S. Computer Science, AI specialization In progress.
Sophia Learning 30 ACE-recommended college-credit courses Verified on Credly.

AVAILABLE FOR SENIOR AI ROLES.

Roles where product judgment, hands-on implementation, quality gates, and release discipline matter. Portland, OR Metro Area. Remote or hybrid.

Or email directly: dbhavery@gmail.com

EMAIL GITHUB LINKEDIN DBHAVERY.DEV