TeaBlendAI – AI-powered Tea Auction Platform

I developed an AI-powered conversational assistant that allows users to interact with TeaBlendAI using natural language for business analytics, Ceylon tea knowledge, data visualization, and authorized auction operations. The chatbot supports multi-turn conversations and can generate dynamic charts, tables, and interactive auction actions.

I built the platform using Next.js, FastAPI, LangChain, Google Gemini, and Model Context Protocol (MCP). I designed and implemented four specialized MCP servers, each responsible for a specific capability: • tea_database — Converts natural-language business questions into T-SQL queries, validates them using read-only SQL safety rules, executes them against the database, and returns structured analytical data. • tea_search — Performs web-based Ceylon tea research using Tavily and Gemini, allowing the assistant to answer domain-specific questions about tea regions, grades, production, brewing, and market knowledge. • tea_visualization — Transforms analytical datasets into dynamic Chart.js configurations and structured tables, enabling the chatbot to present results as interactive visualizations. • tea_auction — Provides controlled access to auction management operations, including creating, retrieving, updating, and deleting auctions while enforcing seller authorization and validation. Beyond the MCP layer, I implemented natural-language-to-SQL generation, intent classification, parameter extraction, multi-turn slot filling, conversation-state persistence, timezone-aware date/time parsing, and AI-generated auction descriptions. The frontend provides a rich conversational experience with interactive charts, tables, auction cards, confirmation actions, dynamic form inputs, conversation search, and PNG/CSV export functionality.

I developed a comprehensive analytics platform that provides insights into tea auctions, sales, purchases, blend profitability, and buyer behavior through KPIs, interactive charts, and analytical tables. The dashboard transforms complex tea-industry data into practical business insights for monitoring performance and decision-making.

I built five responsive analytics views using Next.js, TypeScript, Recharts, FastAPI, MSSQL, and DuckDB, connecting frontend dashboards to dedicated backend analytics repositories and REST APIs. I also implemented a Kimball-based OLAP data model, asynchronous snapshot generation, and SWR-style polling with stale-data fallback for reliable and efficient analytics delivery.