Skip to main content

Multi-Channel Conversational AI

The core chatbot service behind a travel website and WhatsApp bot. It searches flights, hotels, trains and buses, and handles booking workflows through enterprise tool calls.

  • LangChain
  • MCP Tool Calling
  • PostgreSQL + Redis
  • WhatsApp · Web
Illustrative chat showing flight search results with booking buttons
Illustrative chat showing hotel options with an AI review summary

Built at EaseMyTrip. Screens are illustrative mock-ups; internal APIs and data are not shown.

The problem

Travellers ask for things in messy, natural language: "cheap morning flight to Mumbai Friday, window seat". A rigid menu bot cannot handle that, and a plain LLM cannot actually search inventory or make bookings.

The assistant had to understand intent, call the right internal APIs with the right parameters, remember the conversation across turns, and reply in a format that suits the channel, whether that is rich web cards or WhatsApp buttons.

How it works

The architecture, stage by stage. Click any stage to see why it exists.

Dual-chain architecture
01 · Message in

Messages arrive from the website or from WhatsApp via Meta webhook APIs. A router normalises them and loads the user session.

Key engineering decisions

Two chains, two jobs

One chain chooses actions and the other writes responses. Keeping them separate made each easier to prompt, test and improve independently.

LLM Factory pattern

Providers (Groq, OpenAI) can be swapped at runtime, so each request goes to the most cost-effective model that is good enough for it.

Don’t make up parameters

Missing details trigger a clarifying question instead of a guess. A wrong date in a flight search is worse than asking.

AI ⇄ human handoff

State transitions let a conversation move between AI flows, deterministic flows and a live customer-support agent without losing context.

Hard problems along the way

ChallengeThe model sometimes filled tool parameters the user never mentioned.

FixA separate optional-parameter extraction sub-chain with strict “only if stated” rules.

ChallengeThe same answer needs to look different on the web and on WhatsApp.

FixChannel-aware system prompts plus a modular payload builder for text, lists and interactive buttons.

Impact

The service powers the public website chatbot and the WhatsApp travel bot, and is containerised with Docker for production rollout.

Production channels (Web + WhatsApp)
2
Travel verticals via tool calls
4+
Runtime model routing
Groq ⇄ OpenAI