Enterprise AI implementation
HR inbox agent
Built and deployed the agent, worked within approved connections, and refined behavior through example-based evaluation.
Career focus / 03
AI Implementation & Enablement
I’m interested in the part after someone decides, “we should use AI for this.” What should the model actually do? What should stay deterministic? Where does a person need to make the call? And how do we know the whole thing worked instead of just getting a convincing answer back?
That’s the work I’ve been doing professionally with HR agents and automation, and much more extensively at home through Sebastian and my local model work.
Connecting it to real work
Agents, workflows, existing systems, and the information they need.
Deciding what AI shouldn’t own
Identity, permissions, execution, sensitive decisions, and other boundaries that belong in software or with people.
Proving it actually worked
Evaluation, regression tests, release gates, and evidence from the system that owns the result.
Relevant projects
AI-assisted process automation
NCNS case workflow
Used AI for matching, enrichment, and communication inside a larger case workflow where deterministic rules and people retained authority over the consequential decisions.
Status
Tested, approved, and live. Seasonal-volume testing will happen when seasonal volume returns.
Explore projectSystems integration & model evaluation
Sebastian
Built a household AI system across an agent, shared application state, Home Assistant, a media server, physical interfaces, and local models.
Status
Working household system. Model and routing experiments move through separate evaluation stages; not every capability shown in the lab is deployed.
Explore projectImplementation tools and additional experienceExpand for the supporting experience behind these projects.
Implementation details
Access and memory
Some of the system connections I would have liked weren’t available. I worked within the sources and access approved with GRC and system owners, and used Outlook drafts as a persistent memory surface when the agent’s built-in memory wasn’t enough.
Testing the behavior
I reviewed and scored example responses, then changed the instructions, skills, and memory. Requests needing judgment stayed with people.
How I use AI
AI writes a meaningful amount of my implementation code. I define what I need, work through the requirements and architecture, connect the pieces, test the behavior, and keep changing what doesn’t work.
Workflow automation and handoffs
My wider work includes Power Automate workflows, Smartsheet applicant tracking, Microsoft Forms intake, SharePoint records, and Workday reporting.