E-commerce2023· Neurobloom — India/UAE

Neurobloom AI

AI-Powered D2C Wellness Commerce

AI-first Shopify store with conversational commerce, personalised supplement stacks, and cross-border expansion to UAE.

Headline outcome

58% higher AOV · 71% consultation completion · 4-month UAE ROI

An empty operating room, lights and instrument trolleys in place — Illustrative photography. No imagery on this site is from a client environment.

Shopify PlusConversational AIFastAPILLM Fine-tuningHeadlessArabic RTL

Overview

Neurobloom is a D2C wellness brand selling personalised nootropics and supplements. They wanted AI to replace the generic "add to cart" experience with a guided, conversational consultation that recommends personalised supplement stacks.

Challenge

Building a conversational commerce experience on Shopify Plus without a native chatbot framework. The AI recommendation engine needed to consider user health goals, lifestyle factors, and existing supplements to avoid contraindications. UAE market required AED pricing, Arabic UI, and local payment methods.

Solution

We built a custom conversational quiz flow in Shopify Liquid with a FastAPI backend serving LLM-powered recommendation logic. The AI was fine-tuned on nutritional science data and supplement interaction research. UAE expansion included a headless Arabic storefront with regional payment gateway.

Outcome

Average order value increased 58% through AI stack recommendations. 71% of customers complete the consultation flow. UAE expansion achieved ROI in 4 months. 4.8-star average product reviews from AI-matched customers.

Measured

58%

higher average order value

71%

consultation completion

4 months

to cross-border launch in the UAE

The constraint that shaped it

A shopper abandons a conversation faster than a category page, and a recommendation that cannot be explained is worse than no recommendation at all. Both constraints point the same way: the model cannot be the last word. It proposes; the server holds the mapping to what a customer can actually be sold, and the exchange has to resolve in the time a product page would have taken. Flow length, SKU mapping and the second market all followed from that.

How it was built

A consultation instead of a category tree

The shopper never meets a filter panel. The flow asks what they are trying to change, what their days look like and what they already take, then returns a stack with a reason attached to each item rather than a grid sorted by margin. Questions are ordered so the earliest answers do the most narrowing, which is what makes a conversation finishable rather than abandoned halfway: 71% of shoppers who start the consultation complete it. The recommendation set, not a landing page, is the surface being optimised.

The recommendation logic lives on the server

None of the reasoning runs in the browser. The front end collects answers; a FastAPI service holds the fine-tuned model and the mapping from an answer set to real SKUs. That boundary is what makes the system maintainable: the recommendation logic sits behind an interface and can change without a storefront deployment, which is also what stops the consultation and the catalogue from drifting apart.

A second market, not a second build

Cross-border expansion into the UAE completed inside four months, and it read as an expansion rather than a rebuild. The consultation is data the engine reads and not a hard-coded flow, so a second market is a content and commerce exercise carried out around the same recommendation service. The real test of the architecture was that the second market needed no fork of the logic that had made the first one work.

What was hardest to get right

Fine-tuning was not the hard part. The hard part was making a language model behave predictably inside a purchase flow, where an inconsistent answer becomes a refund, a support ticket and a public review. Each turn had to resolve quickly enough that the consultation reads like a page rather than a wait, and the recommended stack had to stand up on its own terms instead of being padded to lift the basket. Average order value rose 58%; the design intent throughout was a larger correct stack rather than an upsell step.

Stack, by concern

Conversation & inference

Conversational AILLM fine-tuningFastAPI

Cross-border commerce

UAE cross-border expansion

More work

GamesTheShop

A custom commerce platform for India's largest gaming retailer

Project SENTINEL

Acoustic Intelligence & Threat Classification

All case studies Start a project