Led design for the company's first chat assistant, as a starting point for other AI initiatives.
2026
Overview
Problem
Equity data is ambiguous — the same terms can be defined differently. Users were navigating calculations across multiple screens and reports, creating risk in high-stakes equity decisions.
My Role
- Led strategy and end-to-end UX for Pulley's first AI product, launching to enterprise customers in June 2026.
- Designed the full chat experience (layout, animations, error flows, and AI personality), set up a continuous dogfooding program, and collaborated with engineering on prompt design to earn a 4/5 satisfaction score across 97 testers.
- Pioneered a "report-first" retrieval flow grounding every answer in verified Pulley data, reducing response time from 90 to 30 seconds with 100% source traceability.
Solution
Pulley AI chat, an assistant that helps make confident, nuanced equity decisions, with every answer sourced from verified company data.
My Team
- Product Manager
- 4 developers (1 front-end, 3 data/AI)
Key Designs
First time experience
Worked on a warm and personal tone to reduce AI use anxiety
Answering with transparency
Every answer traces back to a view in Pulley, so users could verify the math themselves.
Designing for failure
Redesigned error states to fail gracefully, replacing raw error messages with clear follow-ups and reasoning.
Process
V0: Prioritized trust
When I joined, engineering had already shipped an MVP to 40+ companies, and they'd asked over 3,000 questions. I worked with the PM to analyze that data and prioritize my design work. We found that 45% were core equity lookups, like 'who are my top shareholders?', where a wrong answer was costly. UX research made it clear that before we could build ambitious features like AI-written reports, users had to trust the basic answers first. So I focused on trust: adding sources, refining the AI's personality, and reducing error messages.
Analyzed MVP Data
Iterated on Source Designs
Created our voice and tone guide
Pulley had a brand voice guide from an agency, which I evolved into a guide our AI product.
An interesting example where product and brand differed: unlike the brand, which recommended using 'you' or 'I', Pulley AI avoided pronouns. Equity ownership was complicated, and specific names added clarity over pronouns like 'you', or 'she'.
Led testing to cut the error rate
I designed the feedback mechanism so testers could explain why they rated an answer. Across two sprints of company-wide testing, we collected 6,000 ratings, about two-thirds with reasons.
I paired with the error engineer to triage our errors and we cut the error rate by 60%. This cleared us for the customer alpha launch.
Pulley AI