Enterprise architecture. Production AI. Still writing the code.

I build production AI systems inside complex organizations — particularly where the data is sensitive, the infrastructure is constrained, and the systems already in place aren't going anywhere. That means working directly with the people who own the problem, prototyping against their real data, and owning the path from prototype to something that runs unattended.

My work today centres on LLM systems, agents, retrieval, evaluation, model adaptation and the full-stack software around them. Most recently, a six-stage LLM pipeline running entirely offline inside a national laboratory. Behind that sits 25 years of enterprise architecture and delivery — databases, ERP estates, integration, security, migrations — which is usually the difference between an AI prototype and an AI system somebody can actually deploy.

The work maps naturally to forward-deployed engineering: understand the customer's hardest problem, prototype quickly, integrate with the systems already there, and own the path into production.

What I Do

01 — DISCOVER

Find the real problem

I work directly with operators and technical teams to turn an ambiguous operational problem into something that can actually be built — including the uncomfortable part, where the stated requirement isn't the real one.

02 — BUILD

Prototype, then production

LLM applications, agentic workflows, retrieval systems, integrations and full-stack software. Prototype fast against real data, then harden what survives contact with it.

03 — DEPLOY

Ship it and hand it over

Cloud, on-premises or fully air-gapped, with evaluation, observability, security and a handoff plan designed in from the start rather than bolted on when the engagement ends.

Production AI Systems I've Built

06
01

Air-Gapped LLM Migration Pipeline

Sandia National Laboratories · DOE / NNSA
Problem

A federal financial-systems modernization program had to reverse-engineer a large legacy Oracle EBS estate into modern application designs — work that consumed weeks of senior engineering time per application.

Constraint

Air-gapped environment with cloud AI prohibited outright, scarce labelled training data, and a laboratory AI-governance process that had to approve any model before it could be used.

Built

A six-stage offline agentic pipeline — parse, schema mapping, overview, parallel page and inventory generation, assembly — running locally adapted open-weight models behind an automated evaluation gate.

Outcome

Documentation effort cut from weeks to days, 5,000+ lines of implementation-ready design documentation delivered, and no checkpoint scoring under the 90% quality bar ever reached the pipeline. When governance ruled out the original model family, I re-architected onto an approved one without losing pipeline capability.

67%eval-loss reduction
Technical deep dive

XMLTABLE parsing of OAF definitions into a structured intermediate representation; rank-16 LoRA adapters on the attention projections (~15M trainable parameters) trained under both HuggingFace Transformers on MPS bf16 and Apple MLX; Jinja2 slot-based prompt assembly to force deterministic structured output. Training data was the binding constraint, not compute — a scarce set of curated gold examples was expanded 7× with synthetic instruction pairs, and every checkpoint scored against a weighted benchmark (structure, label accuracy, sample-data fidelity, inventory completeness) before promotion. Ran on a Mac Studio M3 Ultra with 256 GB unified memory, packaged for air-gap transfer with reproducible offline installs.

02

Narabi — Production Planning Platform

Global precision manufacturing · Own product
Problem

A global manufacturer ran production planning out of a 700,000-formula Excel workbook — fragile, single-threaded, impossible to query, and understood end-to-end by very few people.

Constraint

Live ERP data, an enterprise security review to survive, minimal client-side footprint, and per-customer isolation as a condition of sale.

Built

A multi-tenant platform: conversational assistant over live production data using retrieval on Oracle 23ai vector search, a containerized ERP integration agent on incremental polls, and a Next.js front end driving a 270-column editable planning matrix with enterprise SSO.

Outcome

The workbook replaced by a real-time platform planners query in natural language. Integration reduced to outbound HTTPS plus a read-only ERP account, which turned security review from a project into a conversation. Weekly release cadence and a documented handoff so the customer owns operations rather than renting them.

700,000formulas replaced
Technical deep dive

QAD ERP synchronized over Progress ODBC on 5–15 minute incremental polls covering inventory, work orders, sales orders and forecasts. Retrieval runs against live planning data rather than a periodically rebuilt index, so the assistant answers from current state. The netting and allocation engine stays in PL/SQL as retained product IP. Per-customer isolated instances on Azure; AG Grid Enterprise carries the editable planning surface at Excel-grade interaction speed.

03

Multi-Agent Document & Supply-Chain Automation

Global manufacturing group
Problem

Invoice and purchase-order processing was handled manually across seven ERP systems spanning three continents, with no consistent path to automate it.

Constraint

Heterogeneous ERPs, multi-region operations, and a genuine accuracy-versus-cost tradeoff across model providers that nobody wanted to bet the process on.

Built

Multi-agent orchestration with policy-based routing, tool invocation and lifecycle observability; document-intelligence agents combining OCR extraction, retrieval and agentic reasoning; and a provider abstraction layer over multiple frontier models with evaluation harnesses measuring accuracy, latency and cost.

Outcome

Document processing automated across seven global ERP systems, with humans placed specifically on the low-confidence tail — where document automation actually fails — rather than reviewing everything. Models can be swapped and A/B tested without touching application code.

7global ERP systems
Technical deep dive

Containerized agents deployed on Oracle Cloud with an event-driven architecture for real-time processing across North America, Europe and Asia. Alongside the document pipeline, agentic supply-chain workflows apply predictive models to demand forecasting, inventory optimization and anomaly detection.

04

Kakoi — Sovereign AI Development Environment

囲い · Own product · kakoi.ai
Problem

Engineering teams in regulated environments cannot send source code to a cloud model, which locks them out of agentic development entirely — the productivity gap is real and growing.

Constraint

Sovereignty over code and data, no software installed on the developer's endpoint, and an audit trail defensible to a security review.

Built

A sealed stack — source control, database, application runtime and a policy-locked agent layer — with provider lock, tool policy, redaction and full audit, behind a single controlled egress gate. Developers connect by web terminal or SSH.

Outcome

Agentic development inside the perimeter, with nothing installed on the endpoint and nothing leaving it unaudited.

0endpoint installations
Technical deep dive

Most of the engineering here is context, not prompting: deterministic context assembly under fail-closed budgets, provenance-tracked retrieval, typed intent classification ahead of the model turn, and immutable graph pinning so agents reason against a single verified view of the codebase rather than drifting mid-session. Grounded change planning, deny-by-default tool policy, and end-to-end auditability throughout.

05

Mitooshi — Sales & Operations Planning

見通し · Own product · mitooshi.io
Problem

Sales and operations planning ran on spreadsheets, leaving no dependable forward view of demand, capacity or raw-material need.

Constraint

It had to share a data platform with the live planning system without ever writing to it, and its numbers had to be provable to planners who would otherwise keep trusting their spreadsheet.

Built

Statistical demand modelling with size and gauge aggregation, finite-capacity modelling, and a rolling twelve-month raw-material forecast, in a separate schema with read-only grants against the planning platform.

Outcome

Backtested to 4.81% MAPE at full coverage before anyone was asked to trust it, with reconciliation gates — tonnage balancing to 0.00 t, money to six decimal places — enforced on every build rather than checked by eye.

4.81%backtested forecast error (MAPE)
Technical deep dive

Requirements were distilled from the customer's own planning spreadsheet — 75 of them — rather than gathered in interviews, which surfaced the rules nobody thinks to state out loud. Every view and page is re-proven live against the repository, so "works on the branch" and "works in the instance" cannot quietly diverge.

06

Agentic Software Engineering

How I build, daily
Problem

Agentic coding tools produce impressive demos and unreliable output at production scale, mostly because the agent is given the wrong context rather than the wrong instructions.

Built

Agentic coding as my primary development environment: custom multi-agent teams for planning, review and security; authored MCP servers for database, application-platform and browser tooling; hooks and skills orchestrating parallel subagents in isolated worktrees. Underneath, a personal memory platform on Oracle 23ai with an HNSW vector index for semantic recall across projects.

Outcome

A working method rather than a demo — retrieval with explicit provenance and supersede semantics instead of append-only notes, context assembly that fails closed rather than silently truncating, and knowledge graphs built at query time so questions are answered against structure rather than flat similarity.

60+projects under semantic recall

How I Work

Most of my career has been spent as the engineer sent into someone else's organization to make something work. That means the technical problem is usually the easier half — the harder half is a stakeholder who describes the requirement incorrectly, a security team that has to be satisfied, a legacy system nobody fully understands any more, and an operations group that will quietly keep using the spreadsheet unless the new thing is genuinely better.

  • Requirements taken from what people actually do, not only from what they say in a workshop — reading the spreadsheet, the production system, the exceptions.
  • Traceability maintained from requirement to delivered screen, so acceptance is a checklist rather than an argument.
  • UAT run as a feedback loop with a real backlog, not a sign-off ceremony at the end.
  • Verification automated before handover — I would rather a harness catch a data problem than have a customer find it in week one.
  • Security and governance treated as design inputs, including federal AI model-approval processes that constrain which models may be used at all.
  • Handover documented so the customer can run and extend the system themselves. Teams I've led have ranged up to eight engineers.

The AI is new. Shipping complicated software into complicated enterprises isn't.

Technical Depth

AI Engineering

LLM application design · agents and multi-agent orchestration · tool calling and MCP · RAG and retrieval architecture · embeddings and vector search · evaluation harnesses and benchmarking · structured output · prompt and context engineering · LoRA/PEFT model adaptation · local inference (Ollama, Apple MLX) · HuggingFace Transformers.

Software Engineering

Python 3.12 with asyncio · TypeScript and JavaScript · Next.js · REST and SOAP API design · system and process API architecture · third-party API integration · AG Grid Enterprise · JSON/XML processing · Git-based delivery.

AI Infrastructure & Deployment

Docker and Docker Compose · Kubernetes · Oracle Cloud and Azure · air-gapped packaging and reproducible offline installs · event-driven architecture · single sign-on and identity (Entra ID) · observability and audit · CI/CD · n8n workflow automation.

Data & Retrieval

Oracle 23ai AI Vector Search · HNSW indexing · expert-level SQL and PL/SQL · data warehousing, star and snowflake schema design · ETL/ELT frameworks · master data management · knowledge graphs · BI Publisher and reporting.

Enterprise Systems

Oracle Database 23ai through 11gR2 · Oracle APEX 24.x back to 5.x · Oracle E-Business Suite R12.2 (Financials, SCM, Projects, HCM, CRM) · ERP integration including QAD · responsibility- and role-based security models · large-scale legacy migration.

Regulated & Constrained Environments

Air-gapped AI operations · federal AI governance and model-approval compliance · national-laboratory and FFRDC delivery · public-trust environments · zero-egress architectures · security review and enterprise threat-surface reduction.

Experience

Since 1998

Sandia National Laboratories · DOE / NNSA

AI-Augmented Financial Systems Modernization

Built the air-gapped local LLM pipeline and fine-tuning toolchain above. Also re-platformed a production non-PO payment application to full functional parity — seven-status approval workflow, four-role security model, legacy-parity audit history — and delivered travel-voucher entry with automatic mileage and GSA lodging proration, plus a reviewer-feedback portal managing UAT across two concurrent modernization projects.

Senior Oracle APEX Architect / Developer · 2026

U.S. Department of Energy / Accenture Federal Services

ERP Modernization, Naval Nuclear Propulsion Program

Technical delivery on the DOE BPMI ERP modernization programme for Bechtel Plant Machinery, the U.S. Department of Energy and the U.S. Navy.

Senior Oracle APEX Architect / Developer · 2022 – 2026

Selected Enterprise Systems Work2014 – 2023
Johnson & Johnson2021–22

Redesigned SMART, the pharmacovigilance platform scanning for statistical drug-safety signals across a multinational user base — roughly 2 GB of new data ingested nightly.

National Oceanic and Atmospheric Administration2020–21

Re-architected all five NOAA Fisheries lending programmes onto Oracle Cloud as eight applications; technical lead for cloud infrastructure and applications.

Federal Aviation Administration2020–23

Designed and built ATOM, consolidating cybersecurity data across FAA systems nationwide into a single view for system owners and auditors.

Estée Lauder — Online Division2019–20

Complete rewrite of the online platform spanning 30+ brands across 80 countries and more than 200 e-commerce sites.

The World Bank — Treasury Division2020

Environment setup and standards for a Treasury-wide platform implementation ahead of migrating 70+ legacy applications.

U.S. Department of Energy — MSTS2017–19

P-Card reconciliation as an ERP extension, plus platform upgrade and ongoing enhancement.

Los Angeles Metro — Office of Management & Budget2015–17

Annual budget forecasting, quarterly amendments and analysis as an ERP extension, with executive performance dashboards.

NASA Jet Propulsion Laboratory2014–15

Technical lead on the Institutional Business Forecast — a 20-year budget forecasting system extending Oracle Financials, integrated with UPS and FedEx logistics APIs.

The Heritage Foundation2023

Requirements through delivery of four applications for a professional networking and resource-sharing platform. Part-time.

Earlier Career1998 – 2014
  • JPMorgan Chase — Mortgage Banking, 78 applications, team of 82012–14
  • Food & Drug Administration — Tobacco Inspection Management2012
  • U.S. Department of Justice — Sentencing Commission, 60,000+ cases2010–12
  • JPMorgan Chase — platform tuning across 700+ applications2010
  • Fairfax County Public Schools — 500-table student data warehouse2009–10
  • Boart Longyear — portal for 4,000+ global drilling projects2009
  • North Carolina DPI — Direct Certification, 18 source systems2008–10
  • Lockheed Martin — U.S. Air Force tooling, 200+ requirements2007–08
  • Getamover.com — CIO, venture-backed startup, grew to 200 staff (acquired)2002–07
  • Navy Marine Corps Intranet — 30 TB security data warehouse, team of 92001
  • The World Bank2000
  • Massachusetts Department of Education2000
  • Baltimore Gas and Electric1999
  • U.S. Department of the Treasury1999
  • United States Postal Service1999
  • Nextel1998
  • AT&T / Lucent1998
  • State of Georgia1998
  • United States Naval Academy1998
Certifications & AwardsSince 1997
Advanced AI for EnterprisesMIT · 2024
Certified Oracle APEX DeveloperOracle · 2015
Certified Tableau AdministratorTableau · 2017
Oracle Top 5 Consultant (award)Oracle · 2000
Oracle Data Warehouse Wizard (award)Oracle · 2000