AI Solutions Architect & UX Product Manager
Overview
Business Objective
After 15 years watching teams skip or shortcut research because of time pressure—then later discover they'd headed in the wrong direction on a hunch—I built an AI agent to fix the problem for myself first. The premise: encode the research questions, source evaluation, and synthesis patterns I'd refined across a career into a tool that takes a team from question to defensible direction in days instead of weeks.
Background
Across consulting engagements, I kept hitting the same friction: clients needed answers fast, research budgets and timelines kept compressing, and existing tools (Dovetail, Notion AI, generic LLM summarizers) didn't weight source credibility or flag where their own confidence should be low. Most produced fluent summaries that read well but couldn't be trusted in client-facing strategic decisions.
Success Measurements
-
Reduce rounds of back-and-forth needed to land on reliable sources and a defensible answer
-
Produce synthesis documentation a non-research teammate can pick up and act on without translation
-
Maintain research rigor (source credibility weighting, confidence flags) while compressing the research timeline
-
Adapt across industries and project types without losing accuracy
Research
The Problem Space
I started by mapping the existing landscape of AI research tools, the workflows I'd been running manually for clients, and the failure modes most common in agent-driven research—hallucinations, opaque sourcing, vendor-funded content treated as neutral. The hardest research challenge wasn't "can the agent find sources"—that's the easy part. It was "can the agent reliably distinguish a credible source from a non-credible one, and weight findings accordingly.
Top findings
-
Generic credibility heuristics ("trust the .edu") break the moment you cross into industries with strong vendor-funded research, gray literature, or genuinely contested evidence
-
Standard LLM outputs are fluent enough to feel trustworthy even when the underlying sources are weak
-
Teams using existing tools often skip source verification because the interface doesn't surface it
-
Confidence calibration—knowing what the agent doesn't know—is more important than raw output quality
Solutions
What the Agent Does
The agent takes a research question—framed loosely or precisely—and runs an end-to-end workflow: question refinement, source identification, credibility weighting, evidence gathering, synthesis writeup, and confidence flagging on weak or contested findings. Outputs are structured for direct handoff to product and engineering teams, so synthesis becomes the start of execution rather than the end of discovery.
Key design decisions
-
Source credibility weighting built in from day one, not retrofitted
-
Explicit confidence flags on weak or contested findings, so teams know what to investigate further
-
Industry-flexible: handles healthcare, life sciences, financial services, and consumer—each with its own source credibility patterns
-
UX-driven interface, built for use under deadline pressure rather than configured for power users only
Real-world use
The agent is in active use across my consultancy. On a recent corporate pharma website rebuild coinciding with multiple new product launches, it widened the competitive reference base by at least 50% within the discovery window while reducing time and effort, putting the design phase on stronger footing from day one.
Responsibilities
Solo design and build. I scoped the architecture, defined the research workflow and source credibility framework, designed the agent's prompts and decision logic, and built the interface for use under deadline pressure—iterating with each new engagement.
Current State & Roadmap
In active use across my consulting practice; I continue to refine the source weighting model with each new engagement. Roadmap includes packaging the agent as a product other teams can use, with industry- specific source libraries and team workflow integrations.
