Services
Industries
Hephon is an AI agency in Dubai (Hephon L.L.C-FZ) that builds AI shopping assistants for supermarkets and grocery delivery in Saudi Arabia, the UAE and Lebanon. Our team built the Saudi-dialect assistant inside the Panda app, and we turned what we learned into a product, Hephon Commerce, which also takes orders on WhatsApp.
A shopper messages the store on WhatsApp the way they would message a person: a list, a recipe, a brand name, often in dialect or Arabizi. An AI assistant searches the store’s live catalog, builds a basket with the right prices and stock, and the order goes through only after the shopper confirms it.
Panda, one of Saudi Arabia’s largest grocery retailers, did not lack customer support. The support it had did not match how people shop. A search for “chicken” in the app returned chicken. A message saying the customer wanted to make chicken biryani for six people returned nothing useful. The intent was there and the catalog was there, but nothing connected the two.
At the same time, help ran through a call centre and a generic web form, with no triage and no self-service. With millions of app users and thousands of interactions a day, every simple question waited for a person.
The brief was to move the simpler interactions into the app itself, in Arabic, with help that knows the catalog and respects how Saudi customers actually talk. Those two requirements, knowing the catalog and handling real dialect, turned out to be the hardest parts of the whole build.
Most grocers in the Gulf and Lebanon face the same gap on WhatsApp, which is where many of their customers already send orders. A message arrives with a list, someone reads it, checks stock by hand and types back a total. It works until volume grows, and then orders wait.
The Panda assistant lives inside the Panda app on Android and iOS. It is a custom build for Panda, not Hephon Commerce, and it is not on WhatsApp. The pieces below are what made it work, and they are the same pieces a WhatsApp ordering assistant needs.
The model does not convert messages into formal Arabic before reading them. It works on the dialect as typed. We built a dataset of real grocery queries from Panda’s anonymised support history, labelled by intent, and tuned the prompt and retrieval against it. Synthetic test prompts were never used.
The assistant never generates product names. It queries a live index and gets back specific items with stock status, price and any current promotion. The model handles the language; the catalog handles the products. That separation is what stops it from recommending products that do not exist.
A shopper asking for “خبز عربي”, “khubz”, “pita” or “Saudi flatbread” gets the same results. The synonym map was built with Panda’s category managers, who know which products customers actually substitute for each other.
A shopper writes that they are making kabsa for six. The assistant identifies the dish, scales the ingredients to six servings, maps each one to an item in the catalog and returns a basket with its reasoning shown: the dish it assumed, the variation, the items it picked and why.
Every reply carries a confidence score. When it drops below the threshold, the assistant stops, tells the customer it is connecting them to someone, and routes the conversation to a human agent with the full context attached.
For retailers who want this on WhatsApp, Hephon Commerce is the product we built afterwards. Shoppers ask in their own words in the retailer’s app, on WhatsApp or by phone; it searches the retailer’s own catalog with prices and stock per branch, turns recipes and lists into a basket, and orders only after the shopper confirms.
Money and stock decisions stay with the shopper and the retailer. The assistant proposes; people confirm.
The basket is a proposal. The shopper sees every item, can edit any of it, and nothing is ordered until they confirm. In early Panda prototypes, silently built baskets were either accepted by customers who did not want them or abandoned by customers who did not trust them. Showing the reasoning fixed that.
Below the threshold the assistant hands over instead of guessing. The human agent sees the question, the assistant’s reasoning, the items it considered and why it stopped, so the customer does not repeat themselves.
Set too high, the assistant passes too much to people. Set too low, it answers things it should not. The threshold is adjusted continuously from the support team’s feedback.
Prices, stock and promotions come from the retailer’s own data. The assistant can only offer what the index returns.
Week 1
Work with Panda’s mobile and operations teams from the first day, not after the model was ready.
Week 3
Integration testing started, with real anonymised customer conversations as test data.
Week 8
Live inside the Panda app on Android and iOS.
Eight weeks was for a custom build at national scale. A retailer starting on Hephon Commerce begins from a working product, so most of the time goes into connecting the catalog, branches and WhatsApp number.
A focused first build usually lands between a few thousand and a few tens of thousands of dollars (AED or SAR equivalent), depending on how many systems it touches and how clean your data is. We scope a fixed price after a short discovery, so you know the number before you commit.
No. Panda’s assistant lives inside the Panda app on Android and iOS. It was designed so the same brain could be connected to other channels without rebuilding it, but what is live today is in the app. WhatsApp ordering is what Hephon Commerce does.
No. Panda’s assistant is a custom build we made for Panda. Hephon Commerce is the product we built afterwards from what that project taught us, for supermarkets and grocers who want the same kind of assistant without a custom project.
Yes. The approach works with the dialect as typed rather than converting it to formal Arabic first. Shoppers write Saudi, Gulf or Lebanese Arabic, Arabizi in Latin letters, English, or a mix such as “ruz basmati 5kg”, and the catalog layer matches all of them to the same products.
It never writes product names itself. It queries the live catalog and can only show items the index returns, with their stock status and price. If the index has nothing that fits, it says so or hands over to a person.
No. It builds a basket and the shopper confirms it. Nothing is ordered on the assistant’s say-so.
Each reply carries a confidence score. Below the threshold it hands the conversation to a person with the full context, including what the shopper asked and which products it considered. At Panda that handoff goes through Genesys Cloud CX.
For a custom build, the range below applies and we fix the price after discovery. Hephon Commerce is priced as a product rather than a project; ask us for a quote for your store count and channels.
Our product for retailers
Hephon Commerce
WhatsApp ordering for supermarkets and grocers: shoppers ask in their own words, the assistant answers from your catalog with prices and stock per branch, and orders only after the shopper confirms.
Tell us what you run today. We will say plainly whether this fits, what we would change for your setup, and what a fixed-price first build looks like.
