HOME / BERKELEY IT

Reducing IT Support Volume with a Smarter Search Experience

A natural language search experience designed to help 45,000+ UC Berkeley students find answers to administrative and enrollment information quickly, reducing routine support requests to the Registrar’s Office.

TEAM

1 PRODUCT MANAGER

2 ENGINEERS

2 DESIGNERS

TIMELINE

JUN 2025 — AUG 2025

TOOLS

FIGMA

USERTESTING

SKILLS

USER RESEARCH

INTERACTION DESIGN

PROTOTYPING

USER TESTING

PROBLEM

Repetitive support requests often bury urgent, time-sensitive student issues, making it harder for the Registrar to prioritize critical cases.

UC Berkeley’s Registrar’s Office handles thousands of student inquiries each year, most of which are routine (enrollment deadlines, record requests, password resets). Without a search experience, students struggle to find answers that already exist in CalCentral and instead open support cases for routine questions, wasting time for both students and staff.

Good Samaritan Shelter's existing site was irresponsive on mobile, required constant internet access, and buried critical resources behind confusing navigation. For 5,000+ residents in Santa. Has received recognition by The Rookies, IDA Awards, and Indigo Awards.

The student is looking for information that exists in CalCentral

→

The student cannot seem to find the answer in CalCentral

→

Student submits a support request to the Registrar

→

Urgent requests get buried among routine support tickets

SOLUTION OVERVIEW

A natural language search that surfaces existing answers so students don’t need to open support tickets for routine questions.

Instead of forcing students to navigate fragmented CalCentral pages or submit support tickets for routine questions, we designed a NLP search experience that helps them surface accurate answers directly from existing institutional information. The system is structured in retrieval-based responses rather than generative AI, ensuring reliability in a high-stakes context.

Conversational Search for Student Support

We replaced traditional keyword-based search with a natural language interface that allows students to ask questions the way they actually think about them. The system interprets intent rather than exact phrasing, improving discoverability of existing CalCentral information.

Conversational Search for Student Support

We replaced traditional keyword-based search with a natural language interface that allows students to ask questions the way they actually think about them. The system interprets intent rather than exact phrasing, improving discoverability of existing CalCentral information.

Search Fallback and Error Handling

We designed error states that actively guide students back to relevant information instead of ending in dead ends. When the system cannot directly resolve a query, it surfaces related suggestions and clarifying prompts to help users refine their intent.

RESEARCH METHODOLOGY

We talked to 30+ Berkeley students and 3 staff workers to understand where the experience was breaking down.

Before designing anything, we ran a mixed-methods research sprint to understand how students and staff actually navigate institutional information. We conducted interviews across two groups, 30 students navigating enrollment, financial aid, and academic records, and 3 Registrar staff members managing incoming support requests. Rather than assuming the problem was technical, we wanted to understand the human behavior underneath it.

8+ student interviews

Conducted with UC Berkeley students across enrollment, financial aid, and academic records queries

35+ survey respondents

Berkeley IT staff and students surveyed on search habits and support ticket patterns

USER RESEARCH

CalCentral student dashboard did not match how students actually search for information.

Across our interviews, one pattern emerged consistently — students phrased questions conversationally ("when's the last day to drop a class?") while CalCentral's navigation used formal institutional language ("Academic Records Requests," "Enrollment Policies"). That terminology gap meant students couldn't find information that already existed in the system, and defaulted to opening support tickets instead. 60% of those tickets were for questions CalCentral already had answers to.

60%

Of incoming Registrar are routine questions about deadlines, links, and basic administrative information.

~40min

Saved per staff member per day if routine inquiries are deflected by AI search

USER PERSONA

Eva isn't confused, she's just asking the wrong system the right questions.

Eva is a third-year UC Berkeley student navigating enrollment decisions, financial aid deadlines, and graduation requirements simultaneously. She's resourceful and self-sufficient, she'd rather find the answer herself than wait two weeks for a Registrar email. She needs a system that meets her where she is, not one that requires her to already know the answer to find it.

I'd much rather find the answer myself than wait two weeks for an update, but if I'm making decisions about my graduation requirements, I need to know the source is absolutely correct, not just something AI made up.

USER PERSONA

Setting safety and scope constraints when designing the search experience.

To ensure reliability in a high-stakes academic context, we defined clear boundaries for the AI search system. The tool is designed to handle informational, institution-specific queries such as enrollment deadlines, policies, and student records, while rejecting or redirecting inappropriate or unsupported requests.

I'd much rather find the answer myself than wait two weeks for an update, but if I'm making decisions about my graduation requirements, I need to know the source is absolutely correct, not just something AI made up.

DEFINING THE AI SEARCH EXPERIENCE

Setting safety and scope constraints when designing the search experience.

To ensure reliability in a high-stakes academic context, we defined clear boundaries for the AI search system. The tool is designed to handle informational, institution-specific queries such as enrollment deadlines, policies, and student records, while rejecting or redirecting inappropriate or unsupported requests.

Generative chatbot

Handles any phrasing

No hallucination risk

Conversational tone

Hallucination risk

Unverifiable sources

Retrieval-Based Search

Natural language input

Verified source data

Sourced results

Traditional Keyword search

Accurate results

No hallucination risk

Requires exact terminology

Root cause unsolved

Generative chatbot

Fully constrained

Fully generative

DEFINING THE AI SEARCH EXPERIENCE

Setting safety and scope constraints when designing the search experience.

To ensure reliability in a high-stakes academic context, we defined clear boundaries for the AI search system. The tool is designed to handle informational, institution-specific queries such as enrollment deadlines, policies, and student records, while rejecting or redirecting inappropriate or unsupported requests.

Generative chatbot

Handles any phrasing

No hallucination risk

Conversational tone

Hallucination risk

Unverifiable sources

Retrieval-Based Search

Natural language input

Verified source data

Sourced results

Traditional Keyword search

Accurate results

No hallucination risk

Requires exact terminology

Root cause unsolved

Generative chatbot

Fully constrained

Fully generative

OUR GOAL

Design a smart search experience that helps students find answers independently, reduces Registrar support load, and maintains accuracy and trust.

IDEATION

Using AI tools like Stitch and Claude to rapidly prototype search interaction states.

We explored different interaction states to understand how students would engage with a smart search experience, including query input, no-result cases, and suggestion-based recovery flows.

OUR GOAL

Design a smart search experience that helps students find answers independently, reduces Registrar support load, and maintains accuracy and trust.

IDEATION

Using AI tools like Stitch and Claude to rapidly prototype search interaction states.

We explored different interaction states to understand how students would engage with a smart search experience, including query input, no-result cases, and suggestion-based recovery flows.

KEY DESIGN DECISIONS

Before designing the interface, we had to answer what interaction model would actually serve students best?

We explored four directions from CalCentral's existing keyword search to contextual modals, browsable FAQ structures, and conversational chat to understand which approach could bridge the gap between how students ask questions and how institutional information is organized.

KEYWORD SEARCH

MODAL/CONTEXTUAL SEARCH

FAQ SECTION

CONVERSATIONAL CHAT

Conversational chat was the only approach that resolved the core problem, answering the question, not just finding the page.

KEY DESIGN DECISIONS

Before designing the interface, we had to answer what interaction model would actually serve students best?

We explored four directions from CalCentral's existing keyword search to contextual modals, browsable FAQ structures, and conversational chat to understand which approach could bridge the gap between how students ask questions and how institutional information is organized.

KEYWORD SEARCH

MODAL/CONTEXTUAL SEARCH

FAQ SECTION

CONVERSATIONAL CHAT

Conversational chat was the only approach that resolved the core problem, answering the question, not just finding the page.

USABILITY TESTING

Over 8 weeks, we prototyped a smart search experience for CalCentral across desktop and mobile.

Before

Services buried below a charity homepage with no clear entry point

Overwhelming list of categories with no filtering or hierarchy

No way to see service status, distance, or availability at a glance

After

Users can discover services through categories or browse "Services Near You"

Has secondary filters

Users can easily scan the name, status and primary actions

USABILITY TESTING

Over 8 weeks, we prototyped a smart search experience for CalCentral across desktop and mobile.

Before

Services buried below a charity homepage with no clear entry point

Overwhelming list of categories with no filtering or hierarchy

No way to see service status, distance, or availability at a glance

After

Users can discover services through categories or browse "Services Near You"

Has secondary filters

Users can easily scan the name, status and primary actions

FINAL PROTOTYPE

Over 8 weeks, we prototyped a smart search experience for CalCentral across desktop and mobile, covering key interaction states including queries, error handling, and recovery flows.

OUTCOME

From scattered support requests to a smart search layer for CalCentral, improving discoverability for 50,000+ UC Berkeley students this coming fall.

The AI search prototype was validated through usability testing and is currently scoped for deployment on CalCentral this semester.

~40%

Reduction in time to find answers during usability testing

60%

Of routine inquiries the search layer is designed to deflect

~45k+

UC Berkeley students to be reached at launch

The team at our end-of-summer social!

REFLECTIONS

Designing AI for institutional settings taught me the importance of transparency, reliability, and trust.

In an academic setting where students depend on AI-generated answers for important decisions, I learned that usability depends not only on capability, but also on how clearly the system communicates its sources, limitations, and failure states.

STILL CURIOUS? CHECK OUT MY OTHER CASE STUDIES

EY Studio+

HANDED OFF 2026

Berkeleytime

Building Berkeleytime’s Design System for 50K+ MAU

SHIPPED 2025