AI Architect
Veteran-OwnedLoren Cossette
Modernization that keeps what matters.
When a county analyst with 36 years of experience retires, her knowledge becomes tests the replacement app has to pass. When Finance is drowning in 350 utility invoices a month, a pipeline catches $2.4M in double-payments. When an agency faces 10,000 inaccessible PDFs before an ADA deadline, 30-second conversions replace 1–4 hours of manual work.
Six production AI platforms. Three county departments. Two clouds. Every one designed so the people whose work it inherits still have a job the day it goes live. Delivered by one accountable principal under reviewed-PR governance, with zero production incidents.
Selected affiliations
- Multnomah County, ORCurrent — AI Architect
- UT AustinAI for Leaders mentor
- Johns Hopkins UniversityAI for Enterprise mentor
- USAAPrior — Chief of Staff to CDAO
- EvolvIQPrior — Principal AI Engineer
- Houston Christian UniversityStrategic AI advisor
- PTR TrainingCorporate AI trainer
Named affiliations only; no endorsement implied.
From the Multnomah County AI Program
Delivery Evidence
Dated record of what shipped
Production milestones, validation improvements, and shipped functionality across the Multnomah County AI Program and other work — dated and traceable to specific case studies. The same auditability standard the program maintains internally.
Impact
Program impact
Independent Validation
Verifiable, not just claimed
Every metric on this site can be traced to a public artifact, an industry-standard validation engine, or an external audit. No testimonials. No hand-waving.
Independent PHP Audit · Xingwu
Cross-Validated File Classification
92.3%
agreement on 33,738 UUIDs
Civic Sentinel's six-bucket classification was independently audited by Xingwu's PHP-based site crawler against the same Drupal inventory. Result: 92.3% direct agreement across 33,738 file UUIDs, with reconciled deltas tracked month-over-month.
Civic Sentinel
axe-core · Industry-Standard WCAG Engine
First-Pass WCAG 2.1 AA Compliance
95%+
first-pass axe-core score
Every A11yReady-converted document is validated through axe-core running real automated checks via Playwright in the production container — the same WCAG 2.1 AA testing engine used by accessibility auditors worldwide. Validation runs before any document reaches human review.
A11yReady
SME Review Queue · Deployed UI
Human-Validated Knowledge Capture
20+
SME-confirmed + audit-locked sessions
Subject Matter Experts confirm captured knowledge through the deployed review interface before it becomes authoritative for downstream agents. 20 SME-confirmed knowledge sessions and 32 tribal-knowledge questions answered in a single day prove the workflow operates end-to-end.
UCR Modernization
Four Areas of Work
How the program runs
Architect & Build Production AI Systems
Multi-agent orchestration, LangGraph pipelines, RAG, Vertex AI Model Garden routing (Claude Opus/Sonnet/Haiku, Gemini) — the systems other people put on a roadmap, I put into production. A11yReady (~185K lines Python, 95%+ WCAG), Civic Sentinel (56K files nightly), and Vantage (multi-tenant agentic testing, 718 locked tests) are live examples.
Capture Knowledge Before It Walks Out
Retiring SMEs take decades of tribal knowledge with them. I build typed knowledge corpora, SME-confirmation workflows, and behavioral test suites that turn institutional memory into executable specs. UCR Modernization captured 704 SME sessions and 208 behavioral Playwright tests.
Modernize Legacy Without Breaking It
Multi-agent testing platforms that index legacy codebases, generate regression suites from captured knowledge, and watch for drift on every commit. Lets the modernized system prove behavioral parity before it ships — and gives leadership confidence to greenlight the rebuild.
Run the Program, Not Just the Build
Stakeholder hubs, dated activity feeds, ROI calculators, validation against independent audits, SME approval queues. I close the loop between technical work and executive trust — so AI initiatives stop dying in pilot purgatory.
About
I'm an AI Architect and hands-on engineer. Sole AI resource for Multnomah County, running 6 production platforms across three departments (DCA, DCHS, DDAP) on two clouds (GCP + Azure) — directing agentic development squads under reviewed-PR, git-worktree-isolated governance with zero production incidents. Shipping weekly with a dated public activity feed.
Multi-agent systems on Vertex AI (Claude, Gemini, GPT). LangGraph pipelines. RAG with pgvector. Behavioral test capture from retiring SMEs. Stakeholder hubs, ROI calculators, and validation against independent audits — plus executive training at UT Austin and Johns Hopkins and peer-reviewed AI research on the side.
Technology
The stack deliveredship
AI / ML
Languages
Full Stack
Cloud / DevOps
Data
Work With Me
Start with a working session
Every engagement begins with a 30-minute working session. You describe what you're trying to accomplish; the firm assesses fit and recommends the shape of the work — whether that's a Discovery, an Architecture & Build Sprint, the flagship Embedded Program Lead engagement, or a referral elsewhere.
