Document Intelligence Platform
A headless platform that reads contracts, invoices and almost any business document, and returns structured data.
{
"type": "invoice",
"vendor": "Acme Ltd",
"total": 1240.00,
"lines": 6
} I design and build AI agents that run in production: document understanding, retrieval and the backend services around them. I also lead the test automation that keeps them dependable.
I'm a software engineer at Genworx, where I build agentic AI systems that run in production: LLM agents and retrieval that turn contracts and invoices into structured data. I also led a team of two that built Genworx's org-wide test automation framework. An AI system is only as useful as it is dependable, so I work on both.
Each case study walks through the problem, the architecture, the decisions and trade-offs, what went wrong, and the result.
A headless platform that reads contracts, invoices and almost any business document, and returns structured data.
{
"type": "invoice",
"vendor": "Acme Ltd",
"total": 1240.00,
"lines": 6
} I led two engineers to build the framework every Genworx team now uses, with shared Playwright libraries and tagging standards.
A backend-for-frontend that routes every request over gRPC to the right microservice, with authorization in one place.
An Android app for a lending business that records money lent to each customer and tracks interest owed.
A small version of the systems I build at work: a tool-using RAG agent that answers only from this site, shows each step and cites its sources.
A demo answers one question well. A production system has to answer the next ten thousand reliably. These are the habits I bring to that work, each one from something I've shipped.
Agents answer from retrieved data, not from memory. Retrieval runs over pgvector, and outputs are structured so the next system can check them.
in practice → Document Intelligence PlatformUnit and end-to-end tests are part of the system, not an afterthought. Shared libraries and tag rules keep every team's tests consistent.
in practice → Test Automation FrameworkWork runs in parallel batches with workers, so one slow document doesn't hold up the rest, and a failed step can be retried on its own.
in practice → 10 documents per batchOne gateway, typed gRPC contracts between services, and authorization in a single place, so each service can change without breaking the others.
in practice → Gateway & Integration LayerFrom the data model to deployment on Google Cloud, I'd rather understand the whole path than hand off a piece of it.
in practice → Lending Ledger App, shipped solo“The interesting part of AI engineering starts after the demo works.”
— what I keep coming back to