01AI Operations Platform
Product Owner / Project Manager A genuinely complex operations platform — built around multiple interdependent workflows, where turning a business request into coordinated action involved several steps, people and decisions.
The idea was to explore how AI could sit inside that process: helping structure requests, support task execution, connect related workflows and reduce some of the manual coordination around day-to-day operations.
The interesting part wasn't simply adding a chatbot to the platform. We were looking at how the workflow itself could become AI-assisted.
ChallengeThe real question wasn't “can we use AI here?” It was: Which parts of the workflow should AI handle, which should it assist with, and where should a person stay in charge?
There were several directions we could take.
One general-purpose agentSimplest to build. Also the most responsibility concentrated in one place.
Multiple specialized agentsClearer responsibilities and better separation of tasks — but more moving parts to coordinate.
AI-assisted, human in the loopLess automated on paper, but gave people the ability to review, guide and take over where judgment mattered.
I worked with engineering to weigh these approaches against reliability, implementation effort, maintenance and the actual needs of the customer. None of them was free — the question was which cost made sense for the value we were trying to create.
Cost considerationWe weren't trying to build the most sophisticated setup. We were trying to find the one that made sense for the customer, the team, the workflow and the budget.
02Multi-Database AI Product Discovery
Product Owner / Project Manager Five separate databases, one product idea: a customer describes what they need once, and the system finds the right product information — wherever it lives. The user shouldn't have to know there are five of anything.
User request→Intent / requirement→Relevant data source(s)→Specialized agent(s)→Product / manual→Answer
ChallengeHiding a genuinely messy data landscape behind one simple conversation.
A — One general-purpose agentFastest to prototype. Also the most likely to get overloaded.
B — Database-specific agentsCleaner per database. Cross-database requests get messy fast.
C — Central routing + specialized agentsMore setup up front — but the direction we explored: read intent, then route.
Engineering weighed in on what mattered most technically — integration effort, how routing would work, what it would take to maintain. My job was connecting that back to what the customer would experience.
Business trade-offThe goal was never the most elegant architecture. It was the one that gave the customer the right answer without costing more to build and maintain than it was worth.
03AI Learning Assistant
Product Owner / Project Manager An early concept for a student-facing assistant, worked through with the engineers who'd build it. The obvious version is a chatbot that answers questions. The more interesting version does more than that.
Answer botJust answers the question.
Content navigatorAnswers, and points to the right material.
Interactive assistantDoes both, and generates follow-up practice questions.
The third option was clearly more useful — it was also more to build. We scoped the MVP around what would move the needle for a learner, not around using AI everywhere we technically could.
MVP decisionCan AI make this meaningfully better for a student? That question mattered more than what was technically possible.