haniseoul

HaniSeoul K-Beauty Shopping Agent: a verifiable shopping assistant on Muse Spark, live on www.haniseoul.com. Problem: A foreign shopper looking at a Korean skincare product wants to know "Is this right for my skin and the ingredients I avoid, and what should I buy instead?" Filters can't answer that, and general chatbots invent ingredients and safety claims. In a Korean store, Korean-language labels make it hard to even identify the product in your hand. What it does: On any product page (no login), the shopper asks in their own language or takes a photo of a product on a store shelf. The agent turns the request into strict conditions (hard avoid, prefer avoid, skin type, product type), checks the product against its printed ingredient list, shows exactly why it is excluded (for example "Fragrance at position #40"), and suggests same-type alternatives that pass, with buy links. Follow-ups such as "exclude essential oils too" show exactly what changed. Key idea: the model proposes, deterministic code decides. Muse Spark understands the request, picks tools and chooses which evidence to show, but it never writes a product fact. Code re-checks every verdict and rebuilds each fact from catalog data, so an answer cannot say what the ingredient list does not. It never judges skin from photos and gives no medical advice. Results: - Golden set of 70 cases (50 shopper questions including 10 non-English and 20 condition-changing follow-ups, plus 20 prompt injections): 69/70 passed, 0 invented products, 0 of 109 shown facts differing from the catalog, 0/20 injections changed a verdict, condition, or recommendation. p50 5.0 s, p95 11.6 s per turn. - Store photos: 25 frames shot in a real Olive Young store. 15/15 in-catalog products ranked first, 5/5 out-of-catalog products and 2/2 look-alike names correctly reported as not found. Built during the hackathon (Oct 27 - Nov 4): the Meta Model API client, agent orchestrator, tools, evaluator, evidence ledger, fact builder, label reader and golden set (all in the public repo), plus the product-page panel and photo input on the live site. Already existed before the event and not part of this submission: the catalog of 3,269 skincare formulations, their ingredient lists, and the HaniSeoul product score. Live demo (deployed Oct 28, 14 languages, no login): https://www.haniseoul.com/kbeauty/products/d-alba-d-alba-first-spray-serum-100ml-ga220816024#ask