Driving 31% higher recommendation click-through rates for a West African platform with 115,000+ annotated product images and local-language reviews.

Client Overview

An e-commerce platform targeting urban youth in West Africa wanted to enhance product recommendations using customer reviews and images. Global recommendation engines failed to account for local slang, cultural preferences, and diverse product visuals, leading to low engagement.

Project Scope

DataLens annotated 75,000+ product images and 40,000 review texts across fashion, electronics, and beauty categories. Annotation tasks included:

• •Image classification for style attributes
• •Aspect-based sentiment analysis
• •OCR for multilingual labels

LLM fine-tuning customized models for hyper-local personalization.

Challenges

• •Informal language variations such as pidgin English in reviews.
• •Overabundance of low-quality user-generated images.
• •Integration with real-time e-commerce APIs.

DataLens Solution

The team used hybrid annotation pipelines combining AI pre-labeling with human oversight for cultural accuracy. For fine-tuning, DataLens engineered prompts simulating shopping dialogues, training LLMs on augmented datasets to generate context-aware suggestions.

Results

• •Improved recommendation click-through rates by 31%.
• •Annotation quality reached 94% precision, with models excelling in 85% of A/B tests.
• •7-week delivery, driving a 18% uplift in average order value for the platform.