TLDR
Built a generative UI agent for a retail analytics company. Users could ask questions and get visual answers instead of navigating a complex dashboard.
The MVP shipped in one week. The client presented it at NRF Singapore 2025 and NRF New York 2026.
The Problem
The client is a retail analytics company serving major brands across India and the US.
They had great data. Trend predictions, market insights, competitive intelligence. The kind of stuff their customers actually need.
But nobody was using it.
Their dashboards were too complex. Users had to click through multiple screens, understand the data structure, and figure out how to get what they needed. Most people just gave up.
The Solution
Build an AI agent that responds with interactive visual UI, not just text.
Instead of navigating dashboards, users ask what they want:
- “What are the top trending colors in the US market?”
- “What’s Brand X doing in Europe right now?”
- “Show me emerging micro-aesthetics in women’s knitwear for FW26”
The agent responds with charts, images, and trend visuals that designers and buyers can use.
This was early 2025, before Generative UI was a common pattern.
What I Built
Week 1: MVP
Got a working prototype up in about a week.
After MVP: Evals
This is where most of the work happened. The team was shipping features, but users weren’t happy with the responses. Classic AI product problem.
I worked with their subject matter experts (designers, retail consultants) to:
- Create a golden dataset of expected Q&A pairs
- Build evaluation pipelines to score agent responses
- Set up a system where they could systematically improve the agent
You can’t improve what you don’t measure.
Ongoing: Advisory
After the initial engagement, I stayed on in an advisory role and helped when the team hit architecture or evaluation problems.
Results
- Presented at NRF Singapore 2025 and NRF New York 2026, retail’s largest event
- Team can now iterate on the agent systematically using evals
What I Learned
Everyone wants to tweak prompts. Building evaluation infrastructure paid off more. Once you can measure quality, improvement stops being random.
I’m not a fashion expert. Working with their SMEs to understand what “good” looks like was essential. The golden dataset we built together was more valuable than any prompt I could write.
The MVP got us to a demo. The longer-term value came from helping the team improve the product after I left the room.
Related
- AI Consulting: More about my consulting work