AI
Five years of shipping AI, and the standards I hold it to.
This page is the capability, not the case studies — those are on Work. Here is what I have put in front of users, what I insist on before it goes there, how I use AI to do the job itself, and where I stop.
Shipped into products
-
UKG · Onboarding & Recruiting · 2024–2025
New Hire Assist and the onboarding agent family
I defined the vision and strategy for onboarding AI agents from market, competitive, and customer analysis, then delivered them. New Hire Assist is a personalized, context-aware assistant using retrieval-augmented generation, NLP, and generative AI to answer pre-boarding and onboarding questions, later extended with voice-enabled guidance for compliance and orientation tasks. Alongside it: a culture agent introducing new hires to their direct and indirect teams off position-management services, and predictive analytics dashboards with AI-generated operational digests drawing on sentiment analysis from 30, 60, and 90-day check-ins.
I worked with ML researchers, engineers, and UX on these rather than around them, and validated the bets through staged proofs of concept and pilots before general availability.
- 52% reduction in employee onboarding inquiries
- 96% response satisfaction
- 30% onboarding task efficiency
- 90% administrator productivity, adopting customers
- 96% / 99% / 74% pilot adoption, satisfaction, KPIs met
-
Company Sage · Legal formation · 2025–2026
Agents and retrieval inside a regulated workflow
Hired as an AI-focused product manager to define and deliver agentic and automation-driven products. I delivered a context-aware onboarding agent that used customer, order, and task context to orchestrate post-purchase formation work through conversational workflows, automated task orchestration, and intelligent notifications. I implemented a retrieval-augmented knowledge assistant for contextual Q&A and automated requirement digests serving both customers and internal administrators. And I automated business-name verification by integrating OpenCorporates and state registry APIs for real-time validation inside the intake experience.
- 60% reduction in task completion time
- 53% reduction in time-to-formation
- CSAT improved
The RAG assistant and the name-availability agent have no adoption or deflection metric I can stand behind, so I am not quoting one for them.
The part I care most about
Governance is a product decision, not a review gate
A confident wrong answer inside legal business formation is a compliance problem. Inside enterprise HR onboarding it is a policy problem. In both cases the interesting question is not which model, it is what the system does when the model is unsure — and that has to be decided while the feature is being scoped.
The standards I set and hold features to: prompt design and prompt governance; hallucination prevention; response validation; citation enforcement so an answer traces to the document it came from; curated knowledge sources rather than an open index; and knowledge governance as an ongoing discipline. Underneath all of it, responsible-AI guardrails meaning traceability, explainability, and compliance controls.
Governance work shows up as incidents that never happened. That makes it hard to put a number against and easy to overstate, so I describe it as a discipline and let the interview test it.
AI as the working method
-
Analysis
Program data, reconciled with AI in the loop
The partner program performance report and ARR analysis were research and data analysis I ran with AI assistance against Snowflake. Database analysis, report generation, and dashboards that would have taken hours to days were finished in hours. The reconciliation framework that established the first true view of that program — joining converted paid intake carts and payment identifiers against attribution records and discount codes — was built the same way, then delivered to the C-suite as a data-integrity audit.
-
Discovery
Feedback synthesis and ICP validation
Partner feedback tracked and synthesized with AI, aligned to the program analysis and to ICP verification work, and turned into a partner elicitation playbook. The feedback dashboard and tracker behind it I built myself with Claude and Claude Cowork.
-
Artifacts
Drafted with AI, edited hard
Vision and strategy documents, roadmaps, PRDs, release notes, user personas, journey maps, and requirements documents. AI drafts them; I own what survives. The craft claim I make about specifications — explicit scope boundaries, a decision log naming each call and its owner, failure-mode analysis, non-functional requirements — is about the thinking, and that part is not delegable.
Time saved runs between a couple of hours and about a week per activity depending on the activity. I give the range rather than an aggregate, because an aggregate would flatter it.
-
Validation
Prototype-first requirements
Working prototypes built on existing system functions with Claude Cowork, Cursor, and UX Pilot, then put in front of real users for requirements validation and cohort feedback before engineering starts a build. It changes the conversation from “does this spec describe what you need” to “is this the thing.”
AI in what I own
Products
Shepherd OS is an AI-driven church management platform, live with real congregations, covering administrative workflows, leadership insights, and member engagement automation. A release-communications agent for PMs and BAs is live and selling. Both were built with Lovable and Claude-generated code on Supabase and Vercel. A third platform is in development and I do not describe it as anything else. See Built.
Client work
Through Prime Solutions Group we take AI workflow and agent-definition engagements — deciding what an agent should do, where a human stays in the loop, and what governance it needs before it touches a customer — and we stand up the AI working method inside client product teams rather than running it for them forever. See Consulting.
What I am not
I am not an ML engineer. I do not train, fine-tune, or evaluate models at the weights level, and I would be the wrong person to hire for that.
I am not an architect and I do not write production code. What I do is build prototypes good enough to settle an argument about requirements, and hold a systems-integration conversation with engineering leads through architecture sign-off.
Stating this plainly is not modesty. Every inflated AI claim collapses in the first technical follow-up, and I would rather be hired for the thing I actually do well.