Emma AI Agent
Designing an AI-powered research assistant for analysts
Media analysts read through hundreds of published articles to understand how a topic is being covered. Most of that work is manual. Emma is an AI agent I designed to handle the reading and structuring part, so analysts can spend their time on judgement instead of collection. I designed it from scratch over four weeks, working with a project manager and two developers.
Role:
UX/UI, Product Designer
domain:
Media
year:
2026
Product:
AI research tool
The Challenge

The agent needed to work at three different depths, from a quick summary to a full research pass. Two rules were fixed by the system itself: date filtering had to be exact rather than interpreted by the AI, and the AI model was chosen automatically based on the selected mode. Both rules had to be visible in the interface instead of hidden behind it.

My Process & Approach

Most of my decisions on this project came from a single question: when the system behaves in a way the user did not choose, how do I make that legible instead of confusing? My process included:

1
Research & Strategy
Analyzed competitor websites to identify usability gaps and market opportunities.
2
Wireframing & UX Design
Crafted an intuitive structure for e-commerce, training enrollment, and stylist applications.
3
Visual Design
Designed a sleek, brand-aligned interface that balanced aesthetics with functionality.
4
Developer Handoff
Provided detailed documentation and design guidelines for the development team working on WordPress.
Understanding how analysts work
I mapped how analysts research a topic today: what they need in five minutes, what they need in full, and at which point they stop trusting an automated answer. The three modes came directly from this. Summary, Overview and Deep Research describe three real situations, not three technical settings, so the choice makes sense before you have used the tool once.
Explaining what the system decides for user
Picking a mode also picks the AI model. That is how the system is built, so the dropdown cannot stay active. I disabled it and added a short note explaining how to unlock it. People accept a rule they can read. They lose confidence in a control that simply stops responding.
Keeping the current state visible
Analysts run several searches in one session, so I put the active mode, sub-mode and date range as tags directly above the results. Anyone can see what an answer covered without opening a panel to check. Sub-Mode only appears once a main mode is selected, which keeps the first decision simple and adds detail only when it becomes relevant.
1
Research & Strategy
Conducted deep dives into Taozen’s mission, audience, and competitors to establish a strong foundation.
2
Wireframing & UX Design
Created a structured sitemap and wireframes for a seamless user journey.
3
Visual Design
Developed a warm, immersive aesthetic that reflected Taozen’s identity.
4
Webflow Development
Built a fully responsive, easy-to-manage site with a CMS for effortless updates.
5
Optimization & Testing
Ensured accessibility, performance, and smooth functionality across all devices.
1
Research & Strategy
Studied The Branch’s brand identity and conducted competitor analysis to understand trends in U.S. political media.
2
Wireframing & UX Design
Created a sitemap and wireframes to organize the content for seamless navigation.
3
Visual Design
Designed a modern yet minimal interface that balanced bold branding with clarity and accessibility.
4
Developer Handoff
Collaborated with developers to ensure smooth animations and interactions aligned with the brand identity.
The Result

I designed Emma from scratch and delivered the full scope: research, user flows, wireframes, the complete UI, and the component kit the product shipped with. The wireframes and final UI were approved on the first iteration. They matched the technical brief and also covered the UX logic that was not defined in it, like what the interface does when a mode locks the model, or when a search runs with no filters set.

The developers implemented the design from my handoff, and analysts were working with the MVP two to three weeks later. It gave them a faster way to move through large volumes of articles, control how deep each analysis goes, and reuse the output directly in their own work.

Reflections
I enjoy work like this: complex enough to require real thinking, and clearly useful once it ships. Analysts needed a modern tool for a job they had been doing manually, and a few weeks after handoff they were using one. Designing something that becomes part of how people actually work is what keeps this interesting to me.
Dive into other projects