Building AI Chatbots That Actually Serve Real Customers

Most AI chatbot demos look impressive. Few survive contact with real customers placing real orders in languages the developer doesn't speak.

Over the past year, I built and shipped a production AI chatbot for JIJO Ventures — a regional FMCG brand — that supports 5 languages (Marathi, English, Gujarati, Hindi, Spanish) and runs entirely on WhatsApp. Customers browse products, place orders, and track deliveries without ever leaving the chat.

The Stack

  • LangChain for orchestration
  • OpenAI GPT-4o-mini and Anthropic Claude as the LLM backbone
  • Qdrant vector database for RAG
  • WhatsApp Meta Business API as the interface

What I Learned

1. RAG needs domain-specific tuning. Generic embeddings fail when your product catalog includes regional FMCG items. Building a training dashboard where admins seed domain knowledge made the difference between a toy and a tool.

2. Multilingual is harder than it looks. Language detection, context switching mid-conversation, and handling code-mixed inputs (Hinglish, Marathi-English) required careful prompt engineering and fallback chains.

3. WhatsApp is a different UX paradigm. No buttons, no carousels (at first). Every interaction is text-first. Designing conversational flows that feel natural while being deterministic enough to process orders is a design challenge as much as an engineering one.

The chatbot now handles real transactions daily. Not a demo — a business tool that replaced manual order-taking for an entire sales channel.

← Back to blog