Building SellSnap AI: Seller Campaigns With an Offline-First Demo Path

I shipped SellSnap AI — turn one product photo into listing imagery, promotional posters, and shareable captions for small online sellers.

AI product tools usually assume a live backend on day one. SellSnap assumes a seller who wants to see the flow before trusting the cloud.

The campaign loop

Photo source → crop → background removal → product details → style / scene → generation → poster editor → captions → project library.

Captions target English, Urdu (RTL, Nastaliq-capable fonts), and Roman Urdu — because the audience is not a Silicon Valley demo reel.

Privilege boundary

Architecture rule: the client never holds AI provider keys, never is the source of truth for purchase verification, and never silently mutates credit balances as if it were the server.

Feature-first Flutter with Riverpod. Mock implementations sit beside real ones and are selected via config (hasBackend) — never by accident. Privileged work goes through the SellSnap backend when connected.

That keeps a stolen APK from becoming a stolen OpenAI bill.

Demo mode is a product, not a stub

In demo mode, processing happens on the phone so the full UX is reviewable offline. Connected providers improve quality later — and the UI says exactly that.

Background removal can run on-device via ONNX / U²-Net (OnnxBackgroundRemovalProvider), with a color-distance fallback in compute when the native runtime isn’t available (tests, unsupported platforms, init failure). Export captures widget boundaries to PNG/JPG for gallery or share.

Demo billing is banner-labeled: no real charges; purchases only simulate entitlement on this device.

Trust copy is part of the architecture

Disclosure isn’t a settings dump:

• What the AI does (cutout, scenes, captions)
• Generated images can drift — verify against the real product before posting
• Photos aren’t used to train models without explicit opt-in

Sellers lose money when a generated image lies. The app should say that before they do.

What I’d still harden

Production activation (backend deploy, store products, provider keys, legal URLs, release signing) is a checklist, not a weekend. Cloud quality will always outpace the on-device demo — keep the gap visible. Moderation belongs on both client and server before scale.

Bet: ship the workflow offline first; upgrade the model second.

If you’re building AI for merchants: make demo mode tell the truth, and keep secrets off the device.