The problem: expertise is rarely ready to use
A sales-training agent needs more than facts about a domain. It needs the decision points an expert uses: which details matter, what to ask next, and how to respond to a particular situation. The initial plan depended on call recordings that were not available.
My role
As Founding Product Manager at VEGA, I set product direction and feature priorities from client requirements, working with engineering on the prototype. My work included intake design, PRDs, API contracts, workflow states, acceptance conditions, and the review process for extracted knowledge.
The decision: redesign the input
I shifted knowledge collection toward structured expert interviews. This made the missing information an explicit part of the workflow: collect what an expert knows, extract decision points, review what is useful, and ask targeted follow-up questions.
I also worked through the handoff between structured JSON extraction and a separately developed agent interface. Defining that contract was necessary to make the knowledge usable beyond the extraction step.
Make the output reviewable
Structured outputs and reviewer-facing model explanations made content easier to inspect and issues easier to locate. The useful question was whether an operator could identify meaningful decision points and explain what was missing.
Operator review surfaced 35 findings. I consolidated them into 10 follow-up requirements for the next iteration, which led to targeted follow-up questions, further expert feedback, and a baseline-search layer for missing domain knowledge.
What changed
In the reported workflow comparison, the share of all extracted decision points that operators judged meaningful rose from 20% to 70%. The number of meaningful extracted points increased from 1 to 15.
The results come from operator review during prototype development and reflect combined changes to knowledge intake, available context, and extraction. The measure tracks useful extracted decision points rather than overall model accuracy.
What I take forward
AI product quality depends on what knowledge enters the system, how that knowledge is represented, and whether people can review the result. Designing that feedback loop is part of defining the product itself.
