lexy-boa/whatsapp-agentic-crm

This project, called Avni, is a customer service assistant that runs on WhatsApp for a fashion retail business. Customers can text or send voice notes, including in South Indian languages like Malayalam, Tamil, Telugu, and Kannada, and Avni searches the product catalog, checks whether items are in stock, looks up past orders, and hands the conversation over to a real person when the situation calls for it, such as a refund or a payment dispute. The published version has all real client details replaced with a fictional demo brand called DemoBoutique. The core of the system is a single call to Claude, Anthropic's AI model, that has access to four tools: one for searching products, one for checking inventory, one for looking up orders, and one for escalating to a human. Rather than running the conversation through several separate AI steps that each had to agree with the others, the model itself decides in one pass which of these tools it needs during a given exchange. The author explains that an earlier version of this system used eleven separate steps across two AI calls, and switching to this single tool using approach cut it down to seven steps and made it simpler to add new capabilities later. Voice notes go through a pipeline that transcribes speech using a Whisper model run through Groq, cleans up transcription errors, and detects which regional dialect is being spoken. Product search uses AI generated embeddings stored in a vector database called Qdrant, while PostgreSQL holds the core business data such as orders and inventory. A dashboard called the Control Room gives the business owner a view of conversations, escalations, and system health. The project runs in Docker containers alongside Postgres, Redis, and Qdrant, and includes a demo catalog so it can be tried locally without a real Shopify store. It ships with over 80 automated tests covering the tool logic, voice pipeline, and privacy safeguards.