At USLBM, my ownership within the procure-to-pay project covered backend technical decisions and the direction of its AI and agentic approaches. The work focuses on reducing manual effort in invoice processing and purchasing reconciliation.

The problem

An invoice, purchase order, and receiving record may describe the same transaction without agreeing neatly. Formats vary, fields can be incomplete, and some discrepancies need a person to resolve them.

Automation needs to fit that reality, including the points where a result needs another look.

My area of ownership

I owned the technical decisions on the backend and helped give the AI and agentic work its direction. That meant thinking through how automated steps would fit together, how results would be checked, and where a case needed human review.

My responsibility was the technical approach within that scope. I focused on how the backend could support the workflow reliably and how AI could contribute usefully to it.

What made the work challenging

A plausible match isn’t sufficient on its own. The workflow needs to handle uncertainty and leave enough context for someone to understand why a case was processed or referred for review.

The interesting decisions sit around the model as much as inside a call to it: how the process continues, what needs validation, and when it should ask for help.

How P2P led to a shared service

Working on P2P made the need for document extraction especially clear to me. It also showed me that the capability belonged beyond a single workflow: projects dealing with documents would benefit from a common service rather than each building its own extractor.

That led me to propose and independently build the Enterprise Document Extraction Service. P2P was the work that exposed the need; the extraction service was the broader capability I took the initiative to create.

What this work represents

This project reflects my experience taking technical ownership within a larger effort: making backend decisions, setting direction for AI and agentic approaches, and recognizing when a project-specific problem points to a reusable capability.

Specific matching methods, decision rules, prompts, integrations, internal architecture, and business information remain private.