Shoof
An Arabic-first AI shopping assistant for Saudi Arabia — compares electronics across 8+ retailers with full transparency.
Online electronics shopping in Saudi Arabia is broken in specific ways:
A store lists a product as "in stock" — you call to confirm, they say "actually we ran out yesterday." To price-compare a single MacBook, you open 8 tabs: Noon, Amazon.sa, Jarir, Extra, Sharaf DG — and try to remember which had the best price. Reviews are scattered across YouTube, Twitter/X, Snapchat, with no way to tell which are paid placements. And none of these stores understand "ابغى ماك بوك للجامعة ميزانيتي 6000 ريال" — they all do literal keyword matching.
We saw this pain every day. We decided to build the fix.
Shoof is a Telegram bot that understands your request in natural Saudi Arabic, searches across multiple Saudi merchants in real time, and recommends the best option — with full transparency about what commission we earn from each merchant.
Type: "ابغى ايفون 17 برو ماكس 256" — get a comparison from Noon, Amazon.sa, Cartlow, Aleph with prices, availability, condition (new/refurbished), and the affiliate commission we earn (or don't) from each one.
Eight Processing Layers — On Every Message
Hand-built Arabic dictionary that knows ابغى, ميزانيتي, قيمنق, تابي, تمارا — with full normalization: ا/أ/إ, ة/ه, etc. Dialect-first design beats any LLM on turn one.
Converts the natural sentence into structured data: product type, brand, model, budget, use case. This structured representation drives every downstream decision.
Live queries via SerpAPI / Google Shopping KSA across Noon, Amazon.sa, Cartlow, Aleph, Jarir, Extra, Sharaf DG and more. Real-time pricing and availability.
Rejects wrong models, wrong sizes, accessories, and misleading listings. The filtering layer is the difference between a useful result and an embarrassing one.
Claude Sonnet 4.6 decides when to search vs when to ask a clarifying question. '13 or 15 inch?' before burning API credits on the wrong search.
Sonnet 4.5 → Sonnet 4.6 → GPT-4o-mini automatic failover. Outages are part of production — service interruption shouldn't be.
Surfaces the affiliate commission from each merchant. Recommends the cheapest option even when we earn nothing from it. This is our moat.
Every interaction logged to Supabase so we improve the system based on real usage patterns. The bot gets smarter every day it runs.
What We Built It With
Lessons for Building Arabic-Native AI
Dialect first, English second. Hand-designing the Arabic dictionary (with normalization: ا/أ/إ, ة/ه, etc.) mattered more than any LLM choice. Saudis write the way they speak, and the system has to meet them on turn one.
Transparency as a moat. Disclosing commission rates costs us short-term revenue but builds long-term trust. It's a moat competitors can't easily copy without losing their own margin.
Telegram before WhatsApp. 5 minutes to set up vs 4–8 weeks for WhatsApp Business approval — gave us a testable version in a day. We'll add WhatsApp later when the model is validated.
Failover everywhere. Every external integration that matters (LLM, search, Telegram) has a fallback. Outages are part of production; service interruption shouldn't be.
Build Your Own AI Assistant
At Conneqt we design and build AI assistants for e-commerce and service businesses, Arabic-native intelligent search engines, automated content platforms, and analytics systems for Saudi enterprises. Shoof is one of our projects — with the same stack and the same approach, we can build one custom to your domain.