AI Consulting · 2026
UX Insights Assistant
A local, retrieval-augmented research assistant for UX research data
- Company
- Personal Project
- Industry
- AI Tooling / UX Research
- Impact
- Cut research synthesis time significantly while keeping every answer traceable to its source
- Skills
- RAG Architecture, Local LLM Inference, Vector Search, UX Research Synthesis, Python
context & problem
UX research generates a lot of unstructured material — interview transcripts, usability notes, survey exports. Finding patterns across sessions is slow, and generic chatbots either can't access this private data or make things up when they don't know the answer.
my role
Designer & builder — I scoped the problem, designed the retrieval pipeline, and implemented it end to end, all running locally for data privacy.
key decisions
1) Local inference over a cloud API
Why? Research data (interview transcripts, potentially sensitive user quotes) shouldn't leave the machine by default.
Trade-off: Slower inference and more setup effort — vs — full data control and zero per-query cost.
2) Retrieval-Augmented Generation instead of fine-tuning
Why? RAG keeps answers grounded in actual source documents and lets the knowledge base be updated instantly (add a new transcript, it's queryable immediately) without retraining.
Impact: Every answer includes a citation back to the exact transcript passage it came from.
solution highlights
- Local embedding + vector store pipeline for interview transcripts and research notes
- Chat interface with inline source citations
- Chunking strategy tuned for qualitative research data (speaker turns, not fixed token windows)
reflection
What worked well: grounding every answer in a citation made the tool trustworthy enough for actual research use, not just a demo. What I'd improve: better handling of cross-document synthesis (patterns spanning many interviews) rather than single-passage retrieval.