SYDEAR

SnapSell AI is a phone-photo listing studio for small sellers. Problem: Small sellers photograph their product, then stall on the listing: writing a good title and description takes longer than the sale, and AI that fills the gaps invents prices, features, and availability the seller never stated. One early live test of this project invented a store availability claim — that failure shaped the whole design: a listing must never state what neither the photo nor the seller proves. Solution: The seller uploads one photo and optionally adds facts they actually know (brand, variant, size). The system returns a studio-style product image and a listing draft in Thai or English, with every unproven item separated into a "needs confirmation" list instead of being written into the ad. How it works (Meta AI tooling): (1) SAM 3.1 detects the product in the photo (box crop path). (2) Muse Image 1.0 rebuilds the crop as a clean studio image, preserving the real label. (3) Muse Spark 1.3 writes the listing copy from two sources only: text readable in the photo, and seller-provided facts. (4) Evidence-first guardrail (fail-closed): copy is parsed sentence by sentence. Any price, size, nutrition, certification, compatibility, or availability claim whose value is not in the photo or the seller facts is stripped into "needs confirmation". A claim that matches a seller value is kept as consistent — explicitly not as independently verified. Health-benefit wording is never exempt. If the model returns empty or malformed output, the user gets a minimal description plus a warning, never the raw untrusted output. Demo: Video: https://youtu.be/iMgAOikjtME — Live browsable demo (static gallery + web-app UI): https://snapsell-demo.pages.dev/ — Live demo with working upload: https://snapsell.sydear.space/ — Full upload flow, runnable locally from the repo: checkout branch backend-reliability, run python3 demo/server.py, open the page, upload a photo. No API key is stored in the repo. Verified live run (2026-10-29, PR #5 code): Lactasoy carton photo → SAM concept "product" → 1600×1600 studio image → real 284-character listing ("แลคตาซอย Lactasoy สูตรต้นตำรับ Original 500 มล." + carton description) with 4 honest needs-confirmation items. Technical execution: Stack: Python stdlib server (demo/server.py), Meta Model API (SAM 3.1, Muse Spark 1.3, Muse Image 1.0), vanilla JS frontend. Tests on the submitted code: guardrail suite 30/30, audit cases 12/12, language-switch check 40 elements / 0 mismatches, live pipeline verified end to end (not mocked). Repository: https://github.com/Techfrontiers/snapsell-ai — Main: 2ec9d4ff (guardrail + gallery attribution merged). Draft PR #4 (web-app UI, GPT review: PASS WITH SCOPE LIMITS) and Draft PR #5 (backend reliability, GPT review: PASS for F-2/F-1 scope). Both PRs remain Draft: GPT reviews did not authorize merge/deploy. They are linked here as the reviewed code, not as merged main. Impact: Built for sellers anywhere, starting from a firsthand Thai seller problem: one photo should be enough to list honestly. The target user is any small seller on Shopee, Lazada, TikTok Shop, Etsy, eBay, or Shopify who has a phone but no studio, no copywriter, and no tolerance for listings that promise what they cannot prove. No real users or sales metrics yet — that is stated plainly, not padded.