SAVI AI Search (The Polis Center)

Building clarity in a complex data system through human and AI interactions.

Designing AI search for a complex community database.

Role

User Experience Designer

Timeline

June 25 - Present

Team

UX Designers, Data Engineers, PMs, Other Stakeholders

Team

4 Designers (Product, Visual, Motion), 3 PMs & Mentors

SAVI is a community data platform with massive datasets. SAVI AI Search was introduced to make complex data more accessible through natural language queries.

Challenge

Data discovery relies heavily on navigating structured hierarchies in the SAVI website, which requires a degree of technical familiarity, limiting non-expert users from fully leveraging the database.

Exploration

Explored AI-assisted search to reduce discovery friction and support non-expert users too.

My Contributions

  • Supported user tests and observed participant behavior.

  • Synthesized research findings to identify patterns.

  • Evaluated AI outputs for ambiguity.

  • Explored and designed interaction patterns.

  • Collaborated with senior designers, engineers, and managers.

Impact

Initial Approach

  • Aligned on the core role of AI:
    Interpret natural-language intent, surface relevant data, and generate contextual summaries, not just retrieve and display raw data.

  • Defined a clear output structure:
    Concise summary, supporting visualization, and transparent source linkage.

  • Clarity on what AI can/cannot do:
    Assist discovery without replacing structured navigation

  • Set baseline interaction expectations:
    Dynamic prompt suggestions, structured result layout, and error handling states.

Early tests revealed that the concept is really well perceived by the users but it lacks a seamless AI and human interaction:
"Excellent concept, but I am not sure how to exactly work with it."

  1. Participants faced uncertainty around what to ask, and vague queries resulted in AI hallucinations.

Participants often paused, unsure of what to search for exactly.


Participants asked, "Overall, Marion County data."Too broad, AI hallucinated irrelevant results.
AI model-specific query:

Compare Marion County with males, females, blacks, and Hispanics in population with low poverty income​.

Fix: Recommended search prompts to support exploration and reduce ambiguity.

  1. Dead-end error states.

When data wasn't found, AI simply displayed nothing found.
Users hit a wall with no path forward

Fix: Suggest related prompts, preserving the search flow and reducing friction.

Next Steps

The feature remains in its early stage.

Moving forward, the focus is on strengthening clarity and transparency within the interaction through user tests and iterations.

Learnings till now

Working with non-deterministic AI in complex data systems reinforced the importance of systems thinking. Interaction design decisions around interpretation, feedback, and recovery directly shape user trust and perceived reliability.

This exploration deepened my understanding of designing around ambiguity, balancing flexibility with clarity while collaborating across product and technical constraints.

rutikabangera8@gmail.com