About project

The Challenge

As Tawseel’s customer base grew rapidly across the MENA region, traditional keyword search was no longer sufficient to meet shoppers’ needs. Many customers struggled to express their intent in a single search term—particularly in Arabic, where linguistic nuances and contextual phrasing often complicated product discovery. Queries like “What do I need for a beach day?” or “شنو أحتاج لرحلة؟” were difficult for conventional search to interpret without a shopping assistant. This limitation created tangible business challenges:

• Longer browsing times and more complex journeys to find relevant products.
• Lower conversion rates and reduced average order value due to missed opportunities for cross-selling and product bundling.
• High bounce rates and reduced engagement on search sessions.
• Competitive gaps with global e-commerce leaders like Amazon, who were already experimenting with AI-driven shopping assistants.

Technology & Approach

Dedicatted developed a scalable, AWS-native AgenticAI solution. This solution, built on Amazon Bedrock, features a modular Model–Agent–Search (MAS) architecture. Custom MCP (Model Context Protocol) servers orchestrate the entire system.

Key elements included:

• Advanced AI Models: Using Amazon Bedrock models like Claude, Nova, and Titan, the assistant understands natural language, reasons about customer intent, and generates clear, actionable recommendations.
• Custom Orchestration Layer (MCP Servers): Purpose-built to manage conversation agents, handle tool orchestration, and isolate third-party connectors. Our servers specialized in product search, trending analytics, ensuring reliable scaling and detailed observability.
• Multi-Agent Orchestration: The MCP layer coordinates specialized tools to deliver intelligent product discovery:
  • StructuredSearch() → retrieves SKUs via marketplace APIs.
  • FreeTextSearch() → executes semantic and hybrid searches using Amazon OpenSearch Service.
  • WebSearch() → supplements catalog data with curated internet sources when needed.
  • GetAnalytics() → integrates co-bought, co-searched, and trending signals to enrich recommendations and provide explainability.
• Observability & Guardrails: Centralized monitoring of MCP server activity, logging of tool call chains, bilingual prompt evaluation, and fine-tuned rate limits to balance performance and cost.
• Operational Design: Containerized MCP servers deployed in an auto-scaling AWS environment, enabling resilience, version control of agent logic, and rapid iteration without downtime.

This AgenticAI-driven architecture provided Tawseel with a flexible orchestration layer that could evolve to support new AI agents, connectors, and customer experiences, while maintaining enterprise-grade control and reliability.

Implementation plan:

1. MVP Planning & Infra Setup – IAM, Bedrock quotas, OpenSearch provisioning
2. Multi-Agent Architecture – agent flows for product search, analytics, internet lookup
3. Prompt Engineering & Evaluation – side-by-side benchmarking
4. Magento Integration – structured API access to products, bundles, metadata
5. Pilot Launch – partial rollout with monitored CVR & AOV uplift

User Experience & Adoption

The AI-powered shopping assistant was seamlessly integrated into Tawseel’s marketplace, creating a natural, intuitive experience for shoppers:

• Conversational Discovery: Customers ask in Arabic or English (“What do I need for a beach picnic?”), and the agent returns curated bundles with clear reasoning.
• Bilingual Support: Seamless switching between languages without loss of context.
• Speed to Value: Shoppers complete discovery 20–25% faster, saving on average 2 minutes per session.
• Trust & Grounding: By combining semantic RAG search with curated Internet sources, the assistant provides not just products, but context for why they are recommended.

Our Solution

The solution became one of the MECA region’s first large-scale AgenticAI shopping assistants, giving Tawseel a measurable competitive edge.

Business Impact: Positive ROI in Year 1

• A first-year payback with operating costs kept to less than 1% of incremental revenue demonstrates the financial resilience and scalability of the solution.
• The solution establishes multilingual GenAI capabilities that differentiate Tawseel from competitors such as Noon and Amazon.ae.By introducing a reusable AI architecture based on the MAS stack, Tawseel gains a flexible platform to support future applications while keeping operations lean.
• By introducing a reusable AI architecture based on the MAS stack, Tawseel gains a flexible platform to support future applications while keeping operations lean.