Customers buy skincare like they're guessing.
Beauty and skincare buyers ask the same five questions before purchase: "Will this work for oily skin?", "Can I use this with my retinol?", "What's actually in it?", "Free shipping yet?", "Do you have this in another size?". The PDP can't answer all five gracefully. Live chat is expensive and slow at 11pm.
An assistant, not a chatbot.
AISHA reads product context from two sources at once: Shopify's .json endpoint (canonical, structured) and live DOM scraping (current, complete, including pricing, variants, badges). Both run in parallel for full coverage on any page.
Responses stream token-by-token from GPT-4o-mini. The greeting appears in under 100ms — before full page context has even loaded — so the customer never sees a blank, thinking-face widget.
Active-ingredient conflict detection.
Skincare ingredients fight each other. Retinols don't play well with AHAs. Niacinamide has interaction edge cases. A naive assistant happily recommends both. AISHA doesn't.
The conflict layer sits between the model and the recommendation surface: any time AISHA wants to suggest a product, the proposed combination is checked against the customer's existing routine (extracted from cart, browsing context, and conversation). Incompatible recommendations are filtered out before they're shown.
"You're ₹240 away from free shipping."
A real-time progress bar reads the cart total from /cart.js and updates as items go in. When the customer is one product away from free shipping, AISHA can suggest the nudge product itself — informed by the conflict layer, so it doesn't break their routine.
A pre-purchase concierge.
AISHA shifts the conversation off live chat for routine pre-purchase questions, runs 24/7, and recommends products with domain logic baked in — not generic LLM enthusiasm. The conflict layer is the difference between "AI on a storefront" and "AI that actually understands what's being sold."