The Foundry AI systems architecture Waukesha, WI
Erin Brady — Principal Architect

The
Foundry

AI systems architecture for businesses that make and move physical things.

Erin Brady, Principal Architect. Thirty years learning how businesses actually work — inside sales desk to VP of Sales and Marketing in industrial automation distribution. Two years building the software that fixes what's broken, as a solo architect shipping production systems across four industries.

The useful part is the combination. Plenty of people can build. Fewer can tell you which process is worth building for — and which one should just be deleted.

bradymaine207@gmail.com 207-680-0562 LinkedIn Waukesha, Wisconsin

Portfolio

This is home turf. Three decades of sitting with plant managers, purchasing agents, and owners, working out where the process actually snags — it's almost never where the process document says it does.

MARK/READY

LIVE (PILOT)

A B2B ordering portal for industrial panel shops. These shops order wire markers by sending whatever they happen to have: a photo of a handwritten list, a spreadsheet formatted however that customer formats spreadsheets, sometimes a legal pad. Someone on the distributor side retypes all of it, and one wrong character means a panel gets built with the wrong labels. The app takes the photo; the Claude API reads the handwriting and maps inconsistent spreadsheet columns into a clean order schema.

The detail that matters

The part I'm proudest of is not the AI — it's the fallback. When confidence on a parsed part number drops below threshold, the system stops trusting the model and drops to deterministic rules, and the interface says plainly which lines were read by AI and which were not. The people who use this have been burned by software that was confidently wrong, and they'll abandon a tool the first time it costs them a job.

React  ·  TypeScript  ·  Cloudflare Pages Functions  ·  D1 SQL  ·  Anthropic API  ·  Magic-link auth  ·  Scheduled Workers  ·  28-test suite

RiPanel Bridge

SHIPPED (SPRINT 1)

Quoting and file-routing middleware connecting a 3D enclosure configurator to an ERP system and to shop-floor CNC equipment, so a configured enclosure becomes a priced quote and a machine-ready file without anybody rekeying it.

The detail that matters

All pricing is recomputed server-side. Client-supplied data is never trusted, because a browser is not a source of truth for what something costs.

TypeScript  ·  Cloudflare Workers  ·  Hono  ·  D1  ·  React Three Fiber  ·  ERP integration

Rivet

IN DEVELOPMENT (DEVICE-VERIFIED)

A field maintenance app for plant equipment. Scans equipment nameplates with on-device OCR and pins each asset as a persistent spatial anchor using LiDAR, so a maintenance tech can walk a plant and find the same machine record in the same physical spot next time. Paired with Gantry, a per-plant management dashboard, and designed from day one as a multi-plant product rather than a single-site tool.

Swift  ·  Vision OCR  ·  ARKit / LiDAR  ·  On-device speech  ·  WebSocket streaming

SitePulse

WORKING APP

Construction field operations. Turns jobsite photos, voice transcripts, and site conditions into a structured daily report with hazard flags and confidence scoring, so a superintendent gets a usable document instead of a folder of images. Built for a construction contractor — The Foundry's first customer.

The detail that matters

Offline-first with conflict-free sync, because the people filing these reports are standing on a site with no signal.

React Native  ·  Claude API  ·  Structured output  ·  CRDT sync  ·  Offline-first DB

Guardian

DEMO (LIVE)

A customer-facing service-status platform for an independent auto repair specialist: a branded landing experience plus an iOS demo that keeps a customer informed about their vehicle without a phone call. Built as a demonstration of what a service business's digital front door can be.

SwiftUI  ·  Cloudflare Pages
Three projects deliberately stopped

Knowing when not to build

01A distributor bridge where the integration premise turned out to be unsupported by the economics of the channel.
02A purchase-order intake tool where existing products already solved it better than I would.
03A scan-to-quote spike that worked but sat behind a dependency that was not moving.

Each one got validated, written up, and stopped. Saying so early saved everybody money. For anything involving process automation, that judgment matters more than build speed.

Before software

I ran sales and marketing for an industrial automation distributor serving the New England OEM market, growing it from under $3M to over $10M. The work that matters most here is what I did with collaborative robotics.

How that growth actually happened is the part I'd tell you about. I put a cobot behind a bar at Allagash Brewing and had it pour beer for people, and the local news covered it. The coverage was nice but it wasn't the point. The point was that a plant manager who watched a robot hand him a beer stopped thinking of cobots as somebody else's technology. I built a program that put cobots in front of high school students and let them program the arm themselves, which did more for our credibility in that region than any campaign I could have bought.

I also founded the company's systems group from nothing, and rolled Salesforce and ERP-based quoting out across a sales organization full of veteran reps who had no interest in typing anything into a system. That's where I learned the difference between deploying a tool and changing a behavior. The honest measure is whether anybody works differently on Tuesday.

Record
Landed and grew the Doosan Robotics cobot line
Brought a collaborative robot line into the house and built the demand for it from nothing.
Doosan Robotics Distributor of the Year, North America, 2024
Plus Doosan Robotics Platinum Distributor status.
Phoenix Contact Automation Elite Distribution Partner
A competitive designation held by very few distributors.

How I work

Daily drivers

Claude Code and Cursor are my daily drivers. I've also shipped finished applications in Grok, Gemini, and ChatGPT, including two native iOS apps built end to end in Grok. Tool-agnostic by practice rather than preference, with real opinions about where each one is strong and where it quietly makes things worse.

Beyond coding agents

The Anthropic and xAI APIs directly, MCP servers and tool integrations as a normal part of the workflow, local inference through Ollama when data cannot leave the building, ElevenLabs for voice and narration pipelines, and NotebookLM for research synthesis.

What I'm deliberately building next: formal evaluation harnesses with test sets and quality metrics, vector-backed retrieval at production scale, and token and cost telemetry instrumented at the agent level.