To comply with my non-disclosure agreement, I have omitted and obfuscated confidential information in this case study.
Enterprise web
Search it, find it, fix it
Role
Founding Designer
COMPANY
Apiphany AI Company
Figma
Timeline
2024-Present
shipped
Yes
I designed a Search feature for a Stealth startup’s enterprise product, navigating a fast-paced, high-pressure, and ambiguous environment.
what i did
As the founding designer, I wore many hats and owned the entire design process for our product—everything from research, UX/UI design, and visual design to customer discovery and team facilitation. I worked closely with the CEO, the Director of Engineering, a team of 15 AI engineers and developers, and even a freelance graphic designer to bring it all together!
results
New Search improved Task Success Rate by 20% and Query Success Rate by 18%. We secured Series A funding and brought on our first B2B customers—big win for an early-stage product!
Challenge
A good search shouldn’t be noticeable, so why are people noticing it?
As the sole designer on the team, it was up to me to conduct research and comb through existing data to understand which parts of search were working and which parts weren’t.
THE problem
Early customer feedback highlights a 40% dissatisfaction rate with result relevance & slower task completion rates.
Previous design that I did at the early stages of our product. When I created this interface, I wanted to showcase compacted views. In doing so, as I reflect now, I added unnecessary visual clutter to the interface.
Granular problems
Problems with previous user interface
process & reasons
Design a powerful, intuitive Search experience.
To create a best-in-class user experience for the AI Search, I focused on the following goals:
Ensure users can easily and confidently interact with the Search through accessible design.
Create easy to digest and easy to read Search outputs.
Enable continuous learning and improvement of our AI Models.
understanding search
What is an ideal AI Search experience?
Tooltips
Natural LanguageSearch
Query
AI Black box
Input
Result
System understands the intent perfectly & presents to the user with the best results.
Easy to understand search results.
Explanation — A great search experience means users find accurate results seamlessly, with minimal effort. I focused on designing simple affordances to guide users in navigating the product’s features effortlessly.
Search Heuristics & Audit
Tooltips
Approach — I did an audit of 20+ best in-class Enterprise products, detailing everything from search term used, interactions, relevancy, placeholder texts, auto-suggest, auto-complete, number of results, sorting & filtering, error states, and more.
What I learnt — I learnt patterns in search UX like communicating intent clearly with placeholder text, brand voice, variety of auto-suggestions and auto-complete based on context, how filters are presented based on users need, and how AI systems handle typos, and fuzzy matches.
SEARCH COMPONENTS
Creating intuitive Inputs
The Search input is the first interaction a user is going to have with the product. It is important to set the right affordances for users, so they understand what is possible to do.
Autosuggestions for Search based on Issue ID input
Input interactions
Natural language search enables users to ask complex queries but it is sometimes vague, and led to inaccurate results which hurt our user adoption.
Where Natural language could have worked? If this was a feature that we wanted to release for a mature product, we could have taken care of edge case by mentioning to the users that the results could be inaccurate, allowing users to experiment. But since this is our core functionality, we want to produce accurate results.
Visual iterations
Natural language search enables users to ask complex queries but it is sometimes vague, and led to inaccurate results which hurt our user adoption.
Where Natural language could have worked? If this was a feature that we wanted to release for a mature product, we could have taken care of edge case by mentioning to the users that the results could be inaccurate, allowing users to experiment. But since this is our core functionality, we want to produce accurate results.
Streamline user input with Autocomplete and Autosuggestions.
❌
❌
Natural language search
Quick summary with key info like status, stage, ID associated with each open issue.
Option 2
Option 1
Fails when the user input is vague
✅
✅
Predicting user input, but not powerful to predict all user needs.
Natural language search enables users to ask complex queries but it is sometimes vague, and led to inaccurate results which hurt our user adoption.
Where Natural language could have worked? If this was a feature that we wanted to release for a mature product, we could have taken care of edge case by mentioning to the users that the results could be inaccurate, allowing users to experiment. But since this is our core functionality, we want to produce accurate results.
navigating edge case
Empowering control with Pre-filtering
When we first started, our search results were far from perfect. The model struggled to make sense of vague user inputs because our Natural Language Processing (NLP) models were still finding their footing. To tackle this, I designed a solution that introduced pre-filters—essentially, a way to tag data more intelligently and decode user intent upfront. We improved our results and my engineering team loved it!
Tooltips
Prefilters help in understanding users intent & help in information discoverability for inconsistent data.
Previous pop modal based search experience that utilized pre-filters to get better results.
Our data had inconsistency—meant that our users would receive different results for same search queries for different components. Pre-filter helps in bringing transparency for our users, who can now see what data is available for different components before performing the Search.
✅
Initially, inconsistent data across components led to poor search results, a pre-filter helps to tag issues for NLP, improving consistency and outcomes.
final design
Perfecting user intent
As our model capabilities grew, I worked on a few design iterations to perfect the Search intent.
Tooltips
Natural language search with filters for information discoverability.
visual iterations
Triggering Search
❌
✅
✅
500px-800px wide standard top header
Visual iterations for the Search
search components
Clear, scannable search results.
How can we provide search results to set clear expectations and build trust as we are improving our models?
Autosuggestions for Search based on Issue ID input
search results explorations
Designing better data tables
Search query using issue id in the previous design.
final design
User centric data tables with contextual filters
Search query using issue id in the previous design.
final design
User centric data tables with AI Insights
Search query using issue id in the previous design.
search components
Drive continuous AI Model Improvement
How can we continuously refine search outputs?
Default State for User Feedback
design ITERATIONS
Enabling user feedback within Search experience
Feedback icons, positioned where it affects change and hence making it easier for users to interact
Gathering rich feedback in the early phase of our product can allow us to offer better AI outputs in the long term.
errors, warnings, etc.
Getting feedback
Input State
Error State
Polite message at the top asking user to retry
Building trust by sharing alternate option to contact
Icons for quick feedback
120px ht = 240 characters for user feedback!
Input and Error States for User Feedback
KPI's & RESULTS
Success Criteria & Results.
Query Success Rate
14/20 of beta users rated the results from the search feature as “highly relevant” or “perfect match”.
Task Success Rate
Task success rate improved from 20% to 85% after implementing new Search..
Satisfaction Score
Qualitative feedback showed that 8/10 users described the feature as “helpful” or “essential” to their workflow.
SIGNIFICANT SETBACKS
We often struggled to maintain product focus and debated key features!!
When I joined the team, I realized that there is misalignment between members of the team on many matters, often causing confusion, releasing half-baked features or sometimes communication issues between engineering and CEO.
I was uniquely positioned to work closely with the user and business goals, which helped me in presenting ideas that could solve both ends of the problem. I found that the best way for us to navigate the uncertainty was to spend more time discussing what works vs presenting the entire design work in one go.
LEARNINGS & TAKEAWAYS
Product Shipped.
All the first customers experienced the work I did. It's a huge privilege to get ownership of a project that so many people will see and experience.
Takeaways..
Working in a Startup helped me in crafting my self learning, team facilitation and rapid designing skills. Because of a complex system, it was hard to break the mindset of database thinking and move to user focused mindset, but it turned out well in the end after iterations after iterations of experiments.
Nurturing good relationships with our stakeholders, helped me alot when I needed support like getting product feedback before an unexpected customer meeting.
Working with the CEO, I learnt embracing iterative improvement. Instead of aiming for perfection in one go, we met regularly focusing on small, consistent improvements over time.
Do more of..
Be meticulously and almost obsessively detail oriented in my work.
