Architecture and provisioning for AI user interfaces backed by agentic logic. An agent holds context and reaches for data on its own, so every one of those paths is somewhere one tenant's information could surface in another's session. The design aimed at isolation by construction rather than filtering after the fact.

• AI agent design
• Multitenant architecture
• Azure

Chatbots that embed into e-commerce sites to answer product questions, each sandboxed and provisioned on demand rather than left running. Dynamic provisioning keeps infrastructure cost proportional to actual use; the sandbox is what makes it safe to run many of them beside each other.

• AI agent design
• Sandboxing
• E-commerce integration

Integration development in Boomi supporting a global telecom supply chain organization's ERP implementation — the connections moving data between the new ERP and the systems the business already ran on.

• Boomi
• ERP implementation

E-commerce interfaces for distributing software, designed and deployed to handle delivery for customers who buy and lead capture for the ones who aren't ready yet.

• E-commerce
• Web development

Years of textbook order records, cleaned and then visualized so the Center could see which schools and communities its materials were reaching: a national heatmap for geographic distribution, bar charts for slicing by demographic, and line charts for change over time. The cleaning was the bulk of it — duplicate customers, inconsistent names and shifting formats all had to be reconciled before any chart on top of them meant anything.

• R
• RStudio
• SQL

An AI agent that lives in Outlook, processes incoming sales orders, checks order status, and answers questions about products and orders in the reply thread. I owned it from design requirements through development and deployment, and wired it to an MCP server so it reads live ERP data rather than a stale export.

• Model Context Protocol
• Dynamics 365
• Azure

The data layer underneath the agent. MCP servers that expose Dynamics 365 Finance & Operations as callable tools (customers, products, sales orders, inventory and product descriptions), plus Data Management Framework operations for bulk imports and exports of enterprise data. Building both halves meant the tool surface could be shaped around the questions the agent actually needed to answer.

• Model Context Protocol
• Dynamics 365 F&O
• Data Management Framework

End-to-end Sales Orders and Sales Returns workflows built on Celigo, moving order data between a Shopify storefront and Dynamics 365 Business Central without a person in the middle. Returns were the harder half, since they have to reconcile against records that already exist in both systems.

• Celigo (iPaaS)
• Shopify
• Dynamics 365 Business Central

Power Automate flows syncing Azure DevOps work items into Microsoft Planner, taking an existing draft workflow to a fully operational one. The payoff was visibility: task views and progress graphs that reflect reality without anyone maintaining them by hand.

• Power Automate
• Azure DevOps
• Microsoft Planner

AstraZeneca had equipment data that existed but wasn't usable: unstructured, and so invisible to anyone trying to ask a question of it. We built a model that transforms it into accessible formats, letting the business generate insights and build further enhancements on top.

• Data modeling
• Unstructured data extraction
• Python

Replaced paper-based QC logbooks with a digital system, reducing manual errors and accelerating audit processes. In a regulated manufacturing environment the hard part isn't digitizing the record; it's producing a digital version whose audit trail still holds up.

• Quality Control systems
• Requirements gathering

Python tooling written alongside the IT department to help find known vulnerabilities across university systems and support patching them. The scripts ran against infrastructure students and staff were actively using, which made being careful the larger part of the job.

• Python
• Network security