I designed the preview-first interaction model for Mercado Libre’s AI Photo Studio, so sellers could inspect, compare and reverse edits before committing.
S / SITUATION
Moderation could identify photo-policy issues, but sellers still needed to correct them. That journey often left the platform and depended on external editing tools. An AI editor could close the gap, provided people trusted what it would change.
Product images represent the seller’s work and directly influence the listing. Applying an AI result without a clear preview could turn a faster workflow into a loss of control over the product being sold.
T / TASK & OWNERSHIP
I set the UX direction for the trust model, entry points and editor states, collaborating with Product, Engineering and ML. My scope covered the interaction and handoff. The team launched the product after my departure, with the preview-first model retained.
A / ACTIONS & DESIGN DECISIONS
The AI proposes a correction.
The seller evaluates before and after.
Confirm the change or return to the original.
Risk-led exploration
I stress-tested more than 40 edge cases with AI-assisted scenario exploration before committing to the interaction direction: a background cut into the product, an incorrect correction or an interrupted process. The purpose was to expose risks early; this was design exploration, not a substitute for usability testing.
Interaction & trust
The editor presents a before/after comparison before anything is applied. Sellers can inspect the proposal, confirm it or return to the original. I treated reversibility as part of the core interaction, because control had to be visible at the moment people evaluated the edit.
Scope & implementation
I scoped the experience around background, resolution and prohibited text or logo issues, rather than a general-purpose editing suite. I documented the behavior of each state and preserved the same interaction logic across app and web, so the team could implement the model after the design handoff.
R / RESULTS & BUSINESS IMPACT
The team launched Photo Studio after my departure, preserving the preview-first interaction I handed off. The reported post-launch outcome was that 70% of sellers who used the tool improved their listing conversion. That figure describes the share of sellers who saw an improvement; it does not describe the size of a conversion lift.
70% is a reported share of participating sellers, not a 70% conversion uplift. Design and handoff were my scope; launch and commercial results were team outcomes.
For AI that changes someone’s work, trust belongs in the interaction architecture. A visible comparison and a clear way back can matter more than removing one more step.
I would evaluate how clearly different seller groups understand recovery, then test whether the same model holds for batch edits and more complex media.
