Home/Work/WhatsApp sales agent on a live catalog
Retail · AI sales agent · Case study

A WhatsApp agent that answers from the catalog, quotes theright market's price and knows when a human shouldtake over.

About 190 variants answered from the live database through direct tool calls, prices from the buyer's market ladder, handover for commissions.

190variants answered from live data
8markets priced correctly
0vector databases
100%of replies traced to queries
WhatsApp · sales agent live
Conversations today14
Handovers2
Cost$0.04
Happening now
  • detect_market(phone)AE4ms
  • get_variants('halo floor')2 results31ms
  • quote_price(H-220, AE)AED 3,9008ms
  • handover(commission)→ Sara
The problem

Stock, price in dirhams, lead time, customisation — a generic chatbot would invent all four. High-value enquiries at all hours, in several countries.

The approach

get_variants, check_stock, quote_price, lead_time against the Prisma catalog; retrieval pipelines evaluated and rejected for this size. Exact answers through direct tools.

Architecture & AI decisions

The first decision was what not to build

A 190-variant catalog does not need a retrieval pipeline; it needs exact answers. Pricing is inherited from the platform's region engine, so the agent quotes the market's own ladder.

01

Groq-hosted Llama 3.3 70B for latency and cost

Groq-hosted Llama 3.3 70B for latency and cost; runtime is model-agnostic

02

Commissions, customisation and discounts hand over to a person with th

Commissions, customisation and discounts hand over to a person with the transcript

03

Trace panel in the back office shows each conversation's tool calls

Trace panel in the back office shows each conversation's tool calls

04

No negotiation

No negotiation: the agent informs, people decide

What shipped.

01

Agent runtime

Model-agnostic tool-calling
02

Tool layer

Live catalog and orders
03

WhatsApp Business API

Production channel
04

Trace panel

Every call visible
05

Handover

Commissions to staff
06

Market pricing

Inherited ladders
07

Guards

No discounts by prompt
08

Content pipeline

Shares the same rules
Result
190variants
8markets
0invented prices
1trace per reply
TypeScriptGroq SDK · Llama 3.3 70BPrismaWhatsApp Business APINext.js back officeVercelTypeScriptGroq SDK · Llama 3.3 70BPrismaWhatsApp Business APINext.js back officeVercel
Selected work

More in production.

All case studies
Healthcare · Conversational AI · SaaS

An AI receptionist a clinic can configure in an afternoon and trust with its calendar.

A prompt compiler builds the agent from onboarding answers; a server-side booking guard checks calendar, duration and rules; Flow Studio lets tenants draw routing.

3channels, one agent per tenant
1session to configure an agent
0double-bookings by design
Read the case study
Inbound call · Northgate Dental
2
find_slots(fri, pm)21ms
booking_guard(2:30)
book(2:30) · release(thu)34ms
Retail · Commerce platform

One back office, eight markets, seven currencies — prices set by hand in every one of them.

Per-market price ladders set by hand, a region engine at the request, DDP checkout, entity routing for payments, and a 17-module admin built as one system with roles.

8markets from one codebase
7currencies, never converted
17back-office modules
Read the case study
Back office · Orders
27
#10482 · United Arab EmiratesShipped DDP
#10481 · SingaporePaid
#10480 · United KingdomMade-to-order · day 3 of 12

Have a system
like this in mind?

Send the problem, not a brief. An engineer replies within one business day.

Start a project