§ About
Head of AI R&D at Patra Corporation. Ten years in software engineering, five of them delivering production AI in regulated domains.
// profile
I'm an AI engineering executive and a hands-on builder. At Patra I founded the AI engineering function, took three generations of AI products from zero to production, and grew the engineering organization behind them to a peak of 100 people.
I architected the Agent Platform, which turns one-off AI use cases into governed, reusable building blocks. I also led the accuracy program that scaled document intelligence to 12,000 documents and 400,000+ data points a month, at mid-90s accuracy across 14+ insurance lines and $0.20 per document.
Today I lead a 45-person AI engineering organization, partner with executive leadership on strategy, roadmap, budget, and Responsible AI governance. I still architect the platforms and write production code myself. Before insurance, I spent four and a half years shipping regulated digital banking on BBVA's platforms, first as a consultant and then in house.
// independent project
OpenCatalyst is a local-first agentic AI desktop app, in the same category as Claude Cowork and OpenAI Codex. I'm its architect and sole engineer: a Rust engine, a React and TypeScript console, a CLI, and a Tauri desktop app that runs on the local machine.
The engine runs a streaming agent loop over sandboxed file and shell tools, and it asks for human approval before it acts. Sessions are event-sourced in SQLite, so a task can resume mid-run after the engine restarts.
A multi-provider gateway streams from Anthropic, OpenAI, Gemini, xAI, and local Ollama models, with prompt caching. Its layered router picks the model for each request from spend caps, task-type rules, session affinity, and scores learned from the results people keep. It fails over when a provider goes down. Team fusion, the multi-agent mode, sends one brief to several models at once, has a judge model merge the drafts, and verifies the result in a scratch workspace.
The agent gets real access to the machine, so every part of that access is constrained. Provider keys sit in an encrypted vault, the agent runs under macOS Seatbelt with scoped write access and a network egress allowlist, and every tool call goes through a four-tier permission ladder. MCP plugins over stdio and HTTP let agents reach outside tools inside those limits.
I ran the build agent-first to prove the method. The Rust engine and the TypeScript client meet at a 61-method typed RPC protocol, and the client is generated from the Rust types through JSON Schema. The architecture rules are automated CI gates with live-engine smoke tests. I directed Claude Code and Cursor inside those gates.
// experience
Lead Patra's 45-person AI engineering organization and architected its Agent Platform, a block-based agentic system with deterministic-first execution and portable deployment. Led the accuracy program that raised extraction across 14+ insurance lines from the 80–85% range into the mid-90s, months ahead of schedule, at $0.20 per document against a $1.00 estimate. Drove the shift to AI-native development across the organization: agentic coding tools, evaluation harnesses to judge AI-generated work, and spec-driven templates. Own the AI budget and advise executive leadership on strategy, roadmap, and Responsible AI.
Led four teams (~20 people) to build and deliver Patra AI, scaling to 12,000 documents and 400,000+ data points extracted monthly. Designed its multi-model architecture (NLP for policy parsing, rules for comparison, LLM extraction with RAG for retrieval) and delivered proprietary models on SageMaker with full lifecycle management. Shipped Patra's first agentic AI system, which found and corrected its own extraction errors; customers saw average handling time drop about 30%.
Matured Policy Checking Next Generation from first working application into a governed production system, with model performance review and formal code review standards. Drove ML development with BERT, transformers, and NLP for insurance document extraction, owning the architecture and data pipeline behind Patra's AI products, plus the security and observability groundwork under them.
Founded Patra's AI engineering function in a zero-to-one role. Designed, built, and shipped Policy Checking Next Generation, the company's first AI-capable application and origin of its entire AI product line. Architected it as a modular monolith on AWS, so a small team could ship and evolve it quickly, and recruited the team that became Patra's AI organization.
Engineering lead for BBVA's Online Account Origination platform, owning the credit card service and expanding what customers could do entirely online. Managed two teams totaling 11 engineers and automation developers, shipping new features and integrating BBVA's new technology stack. Designed and deployed Spring Boot and Kubernetes services on AWS EKS, bridging modern front ends with legacy banking systems. Introduced the team's automation framework and an A/B testing framework for validating business ideas.
Front-end lead for BBVA's Online Account Origination platform, managing a team of two engineers and setting the standards the team built on. Built the shared Web Component library, overhauled the build system, and led a driver's license line-detection feature using Hough transforms. Promoted to Team Lead after seven months.
Developed features across BBVA's Online Account Origination and Mobile Banking platforms, spearheading the integration that brought digital account opening (credit cards, CDs, money markets, loans) to iOS and Android for the first time. Shipped the Address Standardization service and a critical fraud detection fix that earned BBVA's Making It Happen Award. Promoted to Senior Software Engineer within ten months.
Built customer-facing features for BBVA's Online Banking platform as an embedded consultant, including the Credit Card Activation application and reusable widgets like Money Transfer. Increased unit test coverage to over 80% of the application's functionality.
Built a classroom management application in PHP and MySQL that let faculty and staff book and manage classrooms. Delivered a major overhaul of clealabama.com and developed screens and features for law.ua.edu's migration to WordPress.
// patent
Positional, ML-driven document extraction with a guided-review feedback loop, the core technology behind Patra's Policy Checking AI. Co-invented with Tony Li and assigned to Patra Corporation.
// education
Georgia Tech in progress
Western Governors University
Western Governors University
// capabilities
// recognition
Leadership Award (peer- and leadership-nominated, for leadership excellence)
Apex Award (outsized, high-impact results)
Director's Award (given annually to one employee)
Living Our Values Award
Making It Happen Award (for a fraud detection fix that protected customer accounts)
Front-page feature on bbva.com
// next
Open to consulting on applied AI, collaborations, or comparing notes on shipping AI in production.