Kavi - AI Shopping Agent

Project Overview

Overview Kavi is a conversational AI shopping assistant built for the Kapruka Agent Challenge, an open competition where 700+ developers submitted over 200 live, working AI agents. Kavi helps users discover products, get gift recommendations, and complete checkout entirely through natural language — in English, Sinhala, or Tamil. The goal of the challenge was simple to state and hard to execute: build a real, live agent on top of Kapruka's Model Context Protocol (MCP) server, not a prototype or mockup. Kavi is the result — a full end-to-end shopping experience powered by conversational AI. The Problem Traditional e-commerce search relies on filters, dropdowns, and category trees that assume the shopper already knows what they're looking for. Gift shopping in particular often starts with a vague idea ("something for my mom's birthday") rather than a specific product. Kavi was designed to close that gap — letting users describe what they want in plain language and have the agent handle search, recommendation, and checkout from there. What It Does Conversational shopping assistant Powered by Google Gemini, Kavi understands natural-language requests, recommends gifts based on context, searches the product catalog, and walks users through checkout — all through chat. Multi-lingual by design Full UI and conversational support in English, Sinhala (සිංහල), and Tamil (தமிழ்). Language context is dynamically mapped into both the Gemini system instructions and the frontend, so the experience is consistent regardless of language. Live, real-time commerce Product search by keyword, category, price range, and availability, rendered through custom cards, detail drawers, and carousels. Cart management updates totals in real time and enforces sensible limits. Checkout can be completed conversationally or through a guided form, capturing sender, recipient, and delivery details. Orders can be tracked by order number after purchase. Personalization that persists User accounts are backed by Supabase Auth (magic link and Google OAuth), with saved delivery details, a persistent list of frequent recipients that auto-syncs during gifting checkouts, and reminders for birthdays and anniversaries that surface tailored gift suggestions before the date arrives. Grounded in real data Every product, category, and location shown to users is pulled live from the Kapruka Model Context Protocol (MCP) server — not cached or mocked data. Tech Stack AI / LLM: Google Gemini Data layer: Kapruka Model Context Protocol (MCP) server Auth & backend: Supabase (magic link + Google OAuth) Frontend: Custom UI with multi-lingual support (English, Sinhala, Tamil) Outcome Kavi was one of 200+ live agents evaluated by judges in the Kapruka Agent Challenge, out of 700+ developers who entered. It didn't place in the Top 20, but it stands as a fully working, deployed product — handling real product search, multi-lingual conversation, checkout, and personalization end to end. What I Learned Building Kavi meant designing a conversational interface that had to work reliably in three languages, wiring a live MCP integration for real-time data rather than static content, and thinking through the full shopping journey — discovery, recommendation, cart, checkout, and post-purchase tracking — rather than a single isolated feature. It was as much a lesson in product thinking as it was in AI engineering.