Kapruka Octo
A conversational AI shopping companion integrating Sri Lanka's largest e-commerce catalog into WhatsApp and the web.
The problem
Traditional e-commerce browsing creates friction at every turn. Searching through deeply nested categories, filling out multi-step delivery forms, and navigating complex checkout pages leads to abandoned carts, particularly on mobile connections.
In Sri Lanka, the vast majority of consumers spend their daily digital time inside messaging apps like WhatsApp rather than downloading dedicated retail applications. I wanted to see if we could remove the checkout friction entirely by bringing the live Kapruka product catalog directly into natural everyday conversations.
Conversational product discovery
Kapruka Octo acts as an intelligent shopping companion capable of handling customer requests from initial product exploration to final payment:
- Plain-language product discovery: Shoppers describe what they want in natural language, such as asking for a birthday gift within a specific budget, and Octo searches the live inventory using Model Context Protocol tools.
- Visual search with OctoLens: Customers can send a photo of any item. Octo uses Gemini multimodal vision to identify the product and retrieve matching catalog listings in real time without storing user images.
- Multi-lingual communication: Designed specifically for Sri Lankan shoppers, supporting English, Sinhala, Tamil, Singlish, and Tanglish with culturally appropriate conversational nuance.
- Conversational checkout: Delivery addresses and recipient details are collected within the chat thread, generating a single one-tap payment link to complete the order.
System architecture and guardrails
Connecting an autonomous language model directly to live commercial inventory requires strict isolation and guardrails:
- Multi-agent orchestration: LangGraph coordinates specialized agents across product search, cart operations, delivery tracking, and general customer support.
- Real-time catalog integration: Product queries and order generation execute through standardized Model Context Protocol servers connecting directly to Kapruka infrastructure.
- Multi-tier safety screening: Inbound messages pass concurrently through regex validation, Google Model Armor, and classification filters to prevent prompt injections and maintain brand safety.
- Cross-channel memory: Customer preferences, cart state, and order histories synchronize across WhatsApp and the desktop web client.
Tech stack
- Models: Google Gemini 2.5 Flash via Vertex AI
- Agent Orchestration: LangGraph and Model Context Protocol
- Backend Services: FastAPI, NestJS WhatsApp gateway, NATS JetStream
- Data and State: PostgreSQL and Redis