iLuxury — An AI-Forward Luxury Resale Operation

iLuxury & Vintage is a Hong Kong house that sells authenticated pre-owned luxury — Chanel, Hermès, Louis Vuitton, Dior and more — live, on location in Japan, Bangkok and beyond, hosted by Iris and Joey to a community of thousands. It isn't a storefront business; it's a live-selling, community-driven one. So I didn't build them a store. I built them an operation: an immersive brand site, a computer-vision intake tool, and a data layer that turns their WhatsApp community into insight.

1 — The Site: a live-selling community hub

The public face is a fully custom, immersive experience at iluxury.carlfung.dev — rebuilt from a generic template into a dark, cinematic "spatial gallery" with Three.js + GSAP + Lenis. A persistent WebGL scene flies the camera through the brand's real sold pieces; kinetic type, a custom cursor, magnetic buttons, an opt-in ambient soundscape, and a Veo-generated brand film set the tone.

Every section serves the real model: The Lives introduces the hosts and the on-location format; Recent Drops is a pinned horizontal corridor of real pieces shown on recent lives; testimonials and a #ILuxuryVintage wall carry real customer photos; and every call-to-action routes to Instagram, Facebook and the WhatsApp groups instead of a checkout.

Crucially, every photo on the site is real — pulled from the brand's own live drops, not stock or AI imagery — and the consignment flow is reframed around how they actually work.

2 — The Intake Tool: Claude vision → spreadsheet

Consignment intake used to mean photographing items with handwritten sticky notes and hand-typing details into spreadsheets. The LuxeVintage Image Analyzer (Next.js on Vercel) replaces that: drag in a batch of product photos and a Claude vision model extracts brand, material, colour, price and customs category for each one — into an editable table that exports straight to Google Sheets in one click.

Images are processed for analysis only and nothing is stored server-side. End-of-day, a stack of 20–30 consignment photos becomes a clean, tax-ready inventory row set in under a minute.

3 — The Data Layer: insight from the live-selling community

The lives happen on Facebook and the buying happens in WhatsApp groups — which means the real signal lives in chat. I built a private analytics layer over that data (reading the WhatsApp bridge via SQLite) that turns thousands of messages and image bursts into operational insight.

A drop detector recognises each live sell-off (a burst of product images from a host) and a claim ledger tracks every item across the groups — what was shown, what got claimed via reactions, and what's paid or shipped. Across the last 30 days that's ~1,500 items auto-tracked with zero manual entry.

On top sits a suite of eleven analytics views — SQL-backed where the question is simple, Claude-backed where it's smart: reply-latency per customer, drifting-relationship detection with an AI re-opener composer, a conversation simulator that predicts how a specific buyer would respond, topic clustering, reaction analytics and more — the kind of CRM intelligence a live-selling business never normally gets.

The result

Three surfaces, one operation — a brand experience people feel, an intake tool that erases hours of data entry, and a data layer that makes a chat-driven business legible. Real photos, real video, real channels, a real model — and AI doing the heavy lifting at every layer.

Tech Stack