Skip to content

AI systems, internal tools, automation, and the infrastructure underneath them. All running in production.

  • Boosting the efficiency of teams, changing how they operate
  • Cutting costs
  • Generating revenue
  • Saving client & company time

Currently automating manufacturing engineering at a 750+ person PCB manufacturer, where I design and ship systems end to end — architecture through deployment.

See the work
Work

Systems in production.

Most of what I’ve built runs inside an ITAR-controlled environment. The boards go into defense and aerospace programs, so nothing leaves the controlled network. That constraint is the interesting part: it’s why these systems were built on-prem, end to end. What’s here is what each one does and what it changed. Every number is measured.

In production · ~280 jobs/week

Manufacturing Data Extraction Pipeline

A vision-language pipeline that reads customer engineering packages, fills in the manufacturing data a person used to type by hand, and shows every value beside the source it came from.

Engineers stopped hunting through drawings. ~117 hours a week went back to the team.

Python · vLLM · vision-language models · Docker · SQL Server · XML integration · isolated-network deployment

98.8%

Extraction success · n = 1,000+ production jobs

30 → 5 min

Entry time per job · Remaining time is source-linked verification

In production · single box, no internet egress

On-Prem LLM Inference Infrastructure

A 122B-parameter MoE (~10B active) running in production on a single NVIDIA DGX Spark, with no internet egress.

A mid-size manufacturer runs 122B-class AI with zero cloud dependency: 91 days of continuous uptime, 47,973 requests, zero server errors. The recovery path was tested rather than discovered — kill the service and it comes back healthy in 136 seconds, unattended.

vLLM · CUDA · quantization (AWQ) · Docker · systemd · Linux/ARM64 · NVIDIA DGX

91 days

Continuous uptime · Zero restarts

In production · 30k → 25k LOC, 704 tests

Legacy System Migration (Perl → Python)

Ported a ~30,000-line legacy Perl application to ~25,000 lines of Python, proved it byte-for-byte identical against the original, and extended it with nine requested features before it replaced the Perl in production.

A system the factory floor depends on daily got modernized without a single day of disruption.

Python · Perl · protocol replay testing · mutation testing · tkinter

704

Automated tests · 0 skipped, from zero

More projects

  • In production · 100K+ contact CRM

    AI Outreach Systems

    Voice and messaging automation over a 100K+ contact CRM for two real estate investment firms, including full carrier compliance.

  • In production · 4 sites

    Multi-Site CAM Automation

    Replaced a manual spreadsheet step with live REST integration into the planning system; rolled out across four manufacturing sites.

  • In progress · CEO-assigned

    Evidence-Cited Planning Co-pilot

    In progress, CEO-assigned. An AI co-pilot that attempts every field, cites its evidence, and flags what changed since last time.

  • Research project · paper trading

    Autonomous Trading Agent

    An LLM-driven options agent built like production software: typed, tested, self-monitoring. (Paper trading — research project.)

Approach

How I work.

Embedded with the people who use it.I sit with the engineers, planners, and clients the tool is for — before it’s built and after it ships — so it fits how they actually work.

End to end.Discovery, architecture, build, deployment, training, and the 2 a.m. failure modes — same person.

Constraints are the interesting part.Mostly on-prem, or through GovCloud — legacy hardware, zero-downtime, the limits usually determine the design.

Toolkit

Toolchain.

Languages
PythonTypeScriptJavaScriptSQLPerl
AI & ML
vLLMmodel quantizationvision-language extractionClaude API / Amazon BedrockRAGprompt architectureagent design
Backend
FastAPIDjangoREST APIswebhooksSQL ServerPostgresSQLite
Frontend
ReactNext.jsReact NativeTailwindVite
Infrastructure
DockersystemdLinuxAWS (S3, GovCloud)VercelSupabaseon-prem GPU
Quality
pytestmypyruffPlaywrightdifferential testingmutation testingCI discipline
Domain
manufacturing systemsCAM automationexport-controlled (ITAR) / isolated-network environmentsCRM & telephony compliance
About

About.

I’m an engineer in San Jose. I studied business analytics at the University of San Diego and taught myself to build, which means I tend to think about the operational problem first and the technology second.

Day to day I’m the sole engineer on AI initiatives at a PCB manufacturer, working the way a forward-deployed engineer does: embedded with the people running production, finding where the work actually breaks, building the system, and staying with it through deployment and iteration. Initiatives come directly from the CEO and the results go back to the executive team.

Nights and weekends I build for clients and for myself.

Open to conversations about hard engineering problems — especially ones where the constraints are unusual.

Contact