Nakhuda - Bilingual AI Shopping Assistant Grounded in the Product Catalogue
AI Shopping Assistant for a Kuwaiti Electronics Retailer
Designed and built a clickable console for an Arabic and English AI shopping assistant at a Kuwait City electronics retailer: every product claim cited to a catalogue field, compatibility decided by fixed rules, nothing added to a cart without the shopper's tap, and a review queue for every suspected invented fact.
Most of our work is under NDA. Per agreement, we can share the work but not the company details - all names and identifying information shown here have been changed.
Impact
Dashboard, playground with tool trace, conversation list, conversation detail and the hallucination flag queue. Each of the 24 demo conversations opens its own detail with claims, tool calls and guardrail events.
Catalogue search, product, compare, stock, delivery, compatibility, cart proposal, warranty, energy label, policy snippet and hand-off. The cart tool only proposes: nothing reaches the cart until the shopper taps Confirm.
Replayed against live release v14 and draft v15 in the playground: VESA fit, stacking kit, price-match refusal, sponsored label, the KM5 refresh rate, fridge clearance, a missing EER and cart confirmation.
7 scripted scenarios in Arabic and English, 14 catalogue products across 8 categories and 14 hallucination flags from the claim checker, shoppers, agents and reviewers.
The Problem
Lulwa Point sells about 14,000 SKUs from six showrooms, a website and an app. Shoppers asked the call centre and showroom staff on WhatsApp the same questions again and again: will this TV fit my bracket, can this dryer stack on my washer, can it reach Jahra before the weekend. Answers depended on who picked up, and a first chatbot pilot invented a refresh rate the catalogue did not list, so it was switched off.
Our Solution
Nakhuda answers shoppers in Arabic or English using only the fields in Lulwa Point's product catalogue, with a source on every claim. Prices, stock and delivery come live from Lulwa Point's systems, compatibility is decided by fixed rules, the cart changes only when the shopper confirms, and merchandisers and reviewers steer and audit it from one console.
The Challenge: An Assistant That Cannot Make Things Up
Lulwa Point wanted an assistant that sells, but that says nothing about a product the catalogue does not say, shows its source for every claim, never fills a cart on its own, and can be steered and audited by the people who answer for it.
Answers From Memory
Specs were quoted from memory by whoever answered the message. A wrong answer about a VESA pattern or a stacking kit meant the return of a 40 kg appliance.
A Pilot That Invented a Spec
A first chatbot built on a generic model stated a TV refresh rate the catalogue did not list. A screenshot went round social media and the pilot was switched off.
Two Languages, One Catalogue
Most shoppers chat in Arabic, right to left, while model numbers, SKUs and VESA patterns stay left to right inside the text. Prices are in KWD with 3 decimals and delivery runs by governorate and area.
Selling Under Consumer Protection
Kuwait's Consumer Protection Law (Law No. 39 of 2014) generally requires accurate product information and prices, no misleading claims and clear warranty terms, so paid placements must be labelled and energy ratings shown exactly as supplied.
Our Approach: From Grounding Rules to a Clickable Console
Spec and Rules Research
Mapped 10 user roles, 31 screens across the shopper chat, hand-off desk and console, and the data model for conversations, messages, claims, tool calls, compatibility rules, cart proposals, guardrail releases, boosts, sponsored campaigns, hand-offs and hallucination flags. Wrote down the rules the design follows: CITRA's Data Privacy Protection Regulation as the privacy baseline, the Consumer Protection Law, GSO energy labelling, Kuwait's 240 V, 50 Hz mains and type G plugs, and AI disclosure as Lulwa Point's own policy.
Clickable Demo on a Production UI Kit
Built 5 console screens on a production dashboard kit. The playground renders the shopper chat as shoppers see it: disclosure card, consent line, product cards with live KWD prices and Sponsored labels, compatibility results, stock and delivery estimates, warranty quotes, energy labels, a comparison with "Not listed" cells and a cart proposal with Confirm and Cancel, beside a tool trace. Arabic chats render right to left. All actions run on local state.
What a Production Build Adds
Integration with a language model (such as Claude via the API or Amazon Bedrock), the PIM (such as Akeneo), the commerce platform for cart and prices (such as commercetools), product search, ERP stock, delivery scheduling, Zendesk or Freshchat hand-off and analytics. The shopper chat in the website and app, the hand-off desk, guardrail releases with approval, merchandising, catalogue grounding, analytics and settings screens, role-based access, an audit log, and review of the privacy and consumer protection wording by Kuwaiti counsel.
Why It Works: Every Claim Has a Source
No Claim Without a Field
Every product fact carries a citation that opens the field, value, unit and catalogue version. The claim checker marks each one supported, unsupported or mismatched, and removes an unsupported claim before the shopper sees it while raising a flag.
Rules Decide, the Model Explains
TV and mount, washer-dryer stacking, fridge space with ventilation clearance, and voltage and plug are checked by fixed rules. The result shows the fields compared side by side, names the accessory needed, and any missing field gives "Can't confirm" with an offer of a person.
The Shopper Stays in Control
The AI badge and Talk to a person are always on screen, prices come from the pricing tool rather than the model, delivery is marked as an estimate, a sponsored card is labelled and never called the best, and every cart change waits for Confirm.
Every Mistake Has a Queue
Flags from the claim checker, shoppers, agents and reviewers land in one queue with severity and age. Reviewers confirm a cause such as model invented or catalogue value wrong, send catalogue errors to the catalogue manager, and pin the case as a regression test.
Related case studies
View all: E-commerce & D2C
Leagăn - Reviews and UGC Platform with Built-In Safety Escalation
Designed and built a clickable reviews and user-generated content platform for a Chișinău baby and nursery retailer selling in Moldova and Romania: a moderation queue that works to a written policy, review detail with personal data masking, and a safety escalation desk that picks up any review describing a hazard to a child.
Tartera - Headless Storefront and Merchandising Admin for Four European Markets
Designed a mobile-first headless storefront for an Andorran outdoor apparel brand selling to Andorra, the EU, the UK and Switzerland, and built a clickable demo of its staff admin: storefront health, merchandising rules, a collection grid editor, CMS content status and per-market prices, taxes and legal settings.
Enterprise Software Engineering - Human + AI
Contact Us
© 2026 - Copyright. All Rights reserved Oceanlab
Location
Global Headquarters
21 Merchant Road, 3rd Floor, Singapore 058267
Main corporate office and leadership operations
North America Office
548 Market Street, PMB 12345, San Francisco, CA 94104
Client operations and business development
India Office
402 Shivalik Shilp II, SG Highway, Ahmedabad 380015
Engineering and delivery center
Enterprise Software Engineering - Human + AI
Contact Us
Locations
Global Headquarters
21 Merchant Road, 3rd Floor, Singapore 058267
Main corporate office and leadership operations
North America Office
548 Market Street, PMB 12345, San Francisco, CA 94104
Client operations and business development
India Office
402 Shivalik Shilp II, SG Highway, Ahmedabad 380015
Engineering and delivery center
© 2026 - Copyright. All Rights reserved Oceanlab