Wowflux: AI-Powered Image Processing Platform for E-Commerce
Context & Problem
Results & Impact
2,000+ images processed in under 5 minutes, scaling to handle increasingly complex sets
Shipped two products 0→1 in 10 months, including TrackMySnap's App Store launch
Live previews and manual override drove adoption among sellers who'd never used AI tools before
Credit-based pricing, introduced ahead of the token/credit model AI products now use as standard
My Role
I joined as Head of Design, running product and marketing end-to-end with an existing ML engine and no product layer. I built and maintained the design system both products drew from, led a team of 4, and partnered directly with data scientists and executives to keep automation aligned with what sellers would actually trust. Startup scope meant owning the full arc: research, IA, interaction design, and visual system, with no handoff between them.
Ideation & Design
Early internal prototypes leaned toward a dense, utilitarian aesthetic — closer to early-2000s tools like WinRAR than anything a seller would recognize. I moved the design language toward something closer to Photoshop: not the feature set, but the visual and interaction vocabulary. Sellers already trusted that interface for image work, so borrowing its familiarity reduced the learning curve for a genuinely new kind of tool: automated editing they'd never used before.
Freeda (desktop)
Built around three core tools — Background Removal, Adaptation, and Combo Generator — each designed to make automation feel reversible rather than absolute.
Background Removal: Started as fully automated, drag-and-drop, no manual controls. User testing surfaced real anxiety about the AI misjudging product edges. We added live previews and single-image test runs before committing to bulk — the fix was visibility, not more automation.
Adaptation: handled resizing/alignment for up to 200 images against marketplace specs. Saveable presets replaced repetitive manual configuration for sellers juggling 50+ spec formats.
Combo Generator: reused the Background Removal upload pattern for familiarity, with a Swap A/B control to reduce alignment errors in merged listings.
TrackMySnap (mobile).
Deliberately not a scaled-down Freeda. Built for sellers on trade floors and warehouses who needed to capture, queue, and download edited images fast — simplicity over feature parity.
Designing for Trust in AI systems
Three principles guided every AI-facing decision on this project, each learned the hard way:
Transparency beats magic. Full automation with no visibility into the AI's reasoning reads as risk, not convenience — users disengage rather than trust a black box. Preview-before-commit patterns convert hesitation into adoption.
Human-in-the-loop scales trust faster than accuracy does. Freeda's manual override wasn't a fallback for AI error — it was the mechanism that let users build confidence incrementally, one job at a time.
Reduce decisions, not just steps. Sellers weren't overwhelmed by using the tool, they were overwhelmed by configuring it. Presets solved a cognitive problem, not a workflow one — worth distinguishing when scoping AI product work.
