Revealing what customers are ready to buy

Every signal was already available. Nothing said where to look.

• Builders have questions, not queries. Working with users, I distilled their goals into key questions - where customers are unhappy, where there’s room, what people pay more for, and who already owns the market.
• Each question became a view that drills in. I mapped each question to the signals that could reliably answer it - then made every view a filter. Select any facet and the whole marketplace narrows with it, so a few clicks can distill thousands of products into a short, relevant list.
• Tuned against real data. Designs were coded against live marketplace data, so real numbers, not mockups, determined which questions held up and which were noise.

Reviews explained the numbers. No one could hold hundreds in their head.

• Searchable wasn’t enough. We made every review searchable first. Working with builders showed they didn’t always know what to ask, or what to make of the results.
• Summaries shaped for product decisions. I structured each app’s review synthesis around what a builder weighs: the jobs customers need done, strengths, weaknesses, and room to compete.
• AI suggests, the builder decides. Every point runs as a similarity search that lists the reviews behind it, ranked by match. The engineers and I worked out the AI pipeline together; I designed and built the interface.

The power was in the controls. So was the learning curve.

• Builders think in goals, not facets. A builder knows they want “apps making $10k a month with bad reviews.” Getting there meant knowing which of dozens of facets to combine, and how. So I let them ask in their own words, and had the product do the translating.
• Each request is also a lesson. A plain-language request sets the same filters people set by hand, and highlights each one, so builders see how their words map to the product. When AI read “$10k a month” as a price instead of revenue, the mistake sat in plain sight, one edit from fixed.
• Built in, not bolted on. Charts, requests, and manual edits all write to one filter state, which I built into the frontend. There’s no separate AI mode to learn; the product works the same however someone starts.