Research Strategy · 2026
AI-Connected UX Research Repository
My role: Product Designer & Research Ops
Tools
NotionClaude (Anthropic API)CursorMixpanelDovetail
Teams
UX ResearchProduct DesignData Analytics
Bottom line
My role: Product Designer & Research Ops. I designed the repository model, the data architecture in Notion, and the AI coding logic end to end.
The problem: UX research kept scattering across tools and studies, with every researcher inventing their own tags, so insights never connected and the repository went stale the moment the next interview landed.
- AI codes every transcript against one shared taxonomy, so new insights connect to prior ones automatically instead of fragmenting.
- The executive summary regenerates on every change, always current.
- Anyone can ask the repository a plain-language question and get an answer grounded in the actual research, with sources cited.
Challenge
Job to be done: a researcher or stakeholder needs to know what's already known about a topic before running new research or making a call, so insights build on each other instead of every study starting from zero.
Qualitative coding is subjective and drifts over time: without a shared taxonomy, every researcher (and every AI pass) labels the same insight differently and the repository fragments.
Keeping databases in sync by hand is unsustainable; re-coding interviews and reconciling analytics is exactly the work that gets skipped first.
Approach
Anchored everything to one taxonomy (Themes, Sub-themes, Meta-themes) so every insight codes against the same spine.
The AI analysis pass reuses existing themes by default, proposing a new one only when content genuinely doesn't fit anything already there.
Kept the AI honest by design: no invented quotes, no confusing speakers, no implying more certainty than the transcript supports.
Built a living executive summary that auto-rolls-up meta-themes and re-versions on every change, plus an 'ask the repository' layer that cites its sources.
Impact
New insights connect to prior ones instead of starting a new vocabulary.
Insights from different studies converge on the same meta-themes, so cross-study patterns surface instead of being lost to inconsistent tagging.
Researchers and stakeholders can interrogate the whole repository in plain language without learning its structure first.
What I'd do differently
Retrospective
- I haven't measured whether this actually cuts time-to-insight for anyone but me, no baseline, no before/after. That's the next thing to prove before calling it a validated system rather than a working one.
- The taxonomy is only as good as how well the AI resists inventing a new theme when an existing one is close enough. I've watched for that failure mode, but haven't stress-tested it against a large, adversarial batch of transcripts.