The Business Challenge
Our client is a boutique luxury diamond brand with showrooms across the Middle East, serving high-net-worth clients who inquire on WhatsApp at all hours. A typical inbound flow: a customer sends a photo of a ring design at 1 AM, asks about pricing in Arabic, and expects an answer before they wake up.
Before Maya, the sales team of six was answering the same 40–50 questions daily — carat sizes, GIA certificates, appointment availability, delivery windows — in English, Arabic, and Hindi. Hot leads went cold overnight. The founder was personally responding to WhatsApp messages during dinner.
They needed a production AI agent that could handle real customer conversations in three languages, read photos of designs and certificates, book showroom visits directly into calendars, and know when to hand off to a human.
Key Pain Points Identified
- Overnight lead decay: 40% of inbound WhatsApp messages arrived between 10 PM and 8 AM — unanswered until morning.
- Language fragmentation: Same product, three languages, six team members with different translation habits.
- Vision problem: Customers sent ring sketches, competitor photos, and GIA certificate images — humans had to visually interpret and respond.
- Calendar chaos: Showroom appointments booked over email, DMs, and phone — double-bookings weekly.
The VisualizExpert Solution
We built Maya — a production WhatsApp AI agent powered by GPT-4 Vision, orchestrated through Make.com, with structured memory in Supabase and calendar handoff to the human sales team.
Our Technical Approach & Architecture
Maya isn’t a chatbot. It’s a full agent with vision, memory, tool use, and escalation rules. Here’s the stack under the hood.
1. WhatsApp Business API via Wati
We connected the client’s verified WhatsApp Business number through Wati, which gave us a webhook for every incoming message — text, image, audio, or document. Every event lands in Make.com within 200ms of the customer hitting send.
2. Vision + Language Understanding with GPT-4
For text messages, we route to GPT-4 Turbo with a system prompt that codifies the brand voice, product catalog, pricing bands, and escalation rules. For images, we call GPT-4 Vision with a specialized prompt that identifies: ring style (solitaire, halo, three-stone), estimated carat range, metal type, and whether the image is a GIA certificate (in which case Maya extracts the report number and looks it up).
Multilingual is handled natively — GPT-4 detects Arabic, English, or Hindi and responds in the same language. Cultural nuances (Arabic honorifics, Hindi formality) are baked into the system prompt.
3. Structured Memory in Supabase
Every conversation, message, and extracted intent lives in Supabase. When a returning customer messages six weeks later, Maya remembers their previous inquiry, budget range, and preferred style. This is what makes it feel like talking to a familiar concierge rather than a fresh chatbot.
conversations ( id, customer_wa, language, last_intent, budget_band, ... )
messages ( id, conversation_id, role, content, vision_extract, created_at )
4. Calendar Handoff & Human Escalation
When a lead qualifies (budget confirmed, timeline within 30 days, specific product interest), Maya offers showroom slots directly from Google Calendar via the Meta API, books the appointment, and pings the human sales team on Slack with a full conversation summary. If Maya is ever uncertain — unusual custom order, negotiation, VIP client — it silently escalates and a human takes over the WhatsApp thread mid-conversation.
The Agent Features
- 24/7 Multilingual Concierge: Handles English, Arabic, and Hindi with cultural awareness — not machine translation.
- Vision-Enabled Product Recognition: Customer sends a photo, Maya identifies the ring style, suggests similar in-stock pieces, and quotes a range.
- GIA Certificate Lookup: Extracts the report number from a certificate image and validates against GIA’s database.
- Booking Automation: Books showroom appointments directly into Google Calendar with buffer time and staff assignment.
- Silent Human Handoff: When a conversation crosses complexity or value thresholds, the sales team is pinged with the full context and can take over mid-thread.
Results & Business Impact
Within the first 90 days of production, lead-to-showroom-booking conversion tripled. The overnight inbound queue — previously 40% of daily volume — now converts at the same rate as daytime, because Maya responds in under 30 seconds regardless of hour.
The founder stopped answering WhatsApp at dinner. The sales team stopped copy-pasting the same answers. And the client’s cost-per-booking dropped 62% because they didn’t need to hire two more sales associates to keep up with growth.
Summary of Wins
Ready to Build an Agent Like Maya?
If your inbound conversations happen on WhatsApp, Instagram, or your website — and your team is drowning in repetitive questions — a production AI agent is faster to ship than most people think. We’ve built 50+ live in production. Let’s scope yours.