I Built a WhatsApp AI Agent for Kochi Metro — Here's How It Works

How I built MetroMind, a WhatsApp-based agentic AI commute assistant
for Kochi Metro using n8n, FastAPI, LangChain, and GTFS data — and placed Top 10
at the OpenAI Codex Nightline Hackathon.

canonical_url: https://midhunpm.in

Real talk — Kochi Metro's app is terrible. You open it, wait for it to load,
tap through four screens, and by that time you've already missed your train.

So I built MetroMind — a WhatsApp AI agent that handles your entire Kochi
Metro commute through a chat. No app switching, no loading screens. Just message
it like you'd message a friend.

It placed Top 10 at the OpenAI Codex Nightline Hackathon. Here's how it works.

---

What MetroMind Does

• Route planning — "How do I get from Aluva to MG Road?" and it tells you exactly which trains, interchanges, and timings
• Fare calculation — real fares pulled from live GTFS data, not hardcoded guesses
• Ticket booking — automated via Playwright since KMRL doesn't have a public booking API (more on this below)
• Natural language — ask it anything in plain English or Malayalam, it figures out what you need

All through WhatsApp. No app install required.

---

The Stack

WhatsApp (Twilio) → n8n → FastAPI → LangChain → GTFS Data
↓
Playwright (booking automation)

• n8n as the orchestration layer — handles the WhatsApp webhook from Twilio, routes messages, manages conversation state
• FastAPI backend — core business logic, GTFS parsing, route calculation
• LangChain — agent layer that interprets user intent and decides which tools to call
• GTFS — Kerala's open transit data format, has all station info, routes, and schedules
• Twilio — WhatsApp Business API for the messaging interface

---

The Interesting Part — Reverse Engineering KMRL's API

KMRL doesn't have a public booking API. Their app encrypts all traffic with AES.

So I reverse engineered it.

Decompiled the APK, traced the encryption keys, figured out the request/response
format. Then built a Playwright-based automation that handles the booking flow
headlessly — it's essentially a bot that fills out the booking form faster than
a human can.

Not the cleanest solution, but it works. And it's way faster than doing it manually.

---

Why n8n Instead of Just FastAPI

I self-host n8n on my homeserver (Dell i5, Ubuntu 24.04, behind Cloudflare Tunnel

• Traefik). Using n8n meant I could:

• Visually wire up the WhatsApp → agent → response pipeline

• Add new triggers and integrations without touching code

• Monitor every workflow run with full execution logs

For a project that's fundamentally about orchestration, a visual orchestration
tool made a lot of sense.

---

What I'd Do Differently

GTFS data freshness — KMRL updates their GTFS feed inconsistently. I ended
up having to manually refresh it. A proper cron job with a diff check would've
been cleaner.

Session management — WhatsApp conversations don't have a native session
concept. I hacked around it with n8n's static data but a proper Redis session
store would've been better.

The AES reverse engineering — fun to do, nightmare to maintain. If KMRL
ever updates their app, the whole booking flow breaks. A proper API partnership
would obviously be ideal.

---

Try It / Source

The project is part of my portfolio at midhunpm.in.

I'm Midhun, a CS undergrad at Sahrdaya College of Engineering building AI agents
and full-stack systems. If you're working on something similar or want to talk
transit tech, AI agents, or self-hosted infrastructure — find me on
LinkedIn or
GitHub.

---

Built with n8n, FastAPI, LangChain, Twilio, Playwright, and way too much
curiosity about how KMRL's app actually works.

Comments

[Dmitry Sergeev]

did you have to deal with a lot of latency using n8n for the whatsapp trigger or is it pretty snappy?

[Midhun P M]

it was pretty laggy and had abit of delay but most of it was due to LLM and the n8n delay was mostly just like 500ms-1500ms max i think 90% of the other delays were just the actual LLM calls and i used a free API so it was kinda slow