Case study
Firework
Three projects where I moved between product strategy, design and engineering to remove friction for three very different users at Firework: creators, shoppers and the sales team. The biggest was an AI system that qualifies and routes over 1,000 leads a week.
Client
Firework
Year
2023 — 2025
Role
UX/UI Designer, Product Manager, Full-Stack Engineer
Platforms
Shoppable video, livestreams, business portal, internal tools
Outcome
An AI lead-qualification system that handles 1,000+ leads a week, plus faster creator onboarding and livestream shopping cues.

Every part of the funnel had its own friction.
Firework helps brands sell through shoppable videos and livestreams. Growth depended on three groups: creators who had to get set up, shoppers who had to decide to buy, and a sales team that had to find the right customers.
Each one was slowed down by something different: hours of re-uploading, no reason to buy now, and days of manual research.
Problem to product, across disciplines.
I worked from problem definition through launch, moving between research, product decisions, interface design and engineering depending on what each project needed.
Creators, shoppers and sales.
Making onboarding effortless for creators, turning livestream viewers into shoppers, and giving the sales team its time back. Each project called for a different skill: product design, product judgment and technical execution.
How much control should creators have over what they import?
3–5 hrs
of manual re-uploading the importer was built to remove
The problem
Creators had spent months building video libraries on Instagram. Joining Firework meant hours of re-uploading, or launching with an empty storefront that looked unprofessional.
The call
Speed versus control. Importing everything is the fastest path to a full storefront, but creators get no say in what appears. Picking posts makes the storefront more intentional, but adds a decision to onboarding. I designed for both: import your whole library in one step, or curate exactly what comes over.
The tradeoff
Offering both meant a more complex interface and more states to build, from the Instagram sign-in and selection grid to live import progress.
What happened
I designed and built the importer for mobile and desktop, including the edge cases most people never see, like failed sign-ins, rate limits and posts that wouldn’t import, so the few creators who hit them weren’t left stuck.
Respect the work people have already done.

How do you encourage people to buy without manipulating them?
8 → 3
urgency ideas mapped by impact and effort, three in the first version
The problem
Before this project, viewers watched livestreams for about 8 minutes on average, but only 2.3% bought. Nothing on screen gave them a reason to buy now, or a sign that anyone else was interested.
The call
I mapped eight ideas by impact and effort, from live purchase counts and limited-stock cues to countdown timers and shareable discount codes. Live momentum and stock cues were the highest impact, and emoji reactions were cheap and made the stream feel alive, so those three made the first version. Countdowns and discount sharing waited.
The tradeoff
Firework couldn’t see completed purchases on brands’ own checkouts, so a live “X people bought this” count would have been a guess. I showed real clicks instead: a softer signal, but an honest one.
What happened
I shipped the notifications as a hypothesis: small cues in the corner that fade after three seconds, visible enough to notice without blocking the video or the products.
Real social proof beats fake scarcity.

What if the sales team never had to research a lead again?
1,000+
leads qualified and routed every week, automatically
The problem
The sales team was spending more time researching customers than selling to them, and different reps judged leads differently, so outreach often went to the wrong companies.
The call
I turned that research into an automated workflow. It pulls company, traffic and tech-stack data, compares each lead to Firework’s existing customers, then asks three AI models (GPT-4o, Gemini and Claude) to qualify it. The majority vote decides, and qualified leads are routed to the right rep.
The tradeoff
Three models cost more than one. My hypothesis was that a majority vote would be more consistent than any single model, worth the extra API calls.
What happened
The system processes over 1,000 leads a week. Caching and running the models in parallel brought each batch from 5 minutes to 30 seconds, about 10x faster, so the sales team could spend its time selling.
Automation isn’t about replacing people. It’s about freeing them to do the work only they can do.

The best solution isn’t always the biggest feature.
At Firework, I learned to approach problems without getting attached to a particular discipline or solution. Sometimes the answer was a better onboarding experience. Sometimes it was a three-second notification. Other times, it was an automated system that removed hours of manual work.
Working across product, design and engineering taught me to think about the whole experience: what users needed, what the business needed and what we could realistically build. The biggest lesson was knowing what mattered enough to build, and what didn’t.
1,000+
leads qualified and routed to the right rep every week, automatically
10×
faster lead batches after caching and parallel models: 5 minutes down to 30 seconds
3–5h
of manual re-uploading the creator importer was built to remove
Video chat, creator tools and AI video.
A 1:1 video chat that shows shoppers who they’re talking to, with an associate’s photo and a live status dot. An animated poster creator that lets anyone make an eye-catching video thumbnail in minutes. And an AI tool that turns a product page into a video ad.
I also merged Livestreams and Digital Showroom into one place to schedule and restream events, and built a guided onboarding for self-serve customers that other teams later reused.
