Bengaluru · since 2019

Abhijit Kumar

Visa Go team · backend

A guest asks where the match is. A spider brings a price back. A bot on my laptop answers from data I hold. Three stories. Visa Go is the one to open.

Stories

One guest. Then the pipes. Then mine.

Visa Go is the story to stay for. The scrape is how the data got trustworthy. The Telegram bot is my own project, a model I run at home.

2024 – now · Visa Go team

A guest has a ticket. The match is somewhere else.

30,000 sponsored guests, handled in the app

Visa sponsored the FIFA World Cup 2026 and the tickets. Visa Go is the invite-only app those guests used to accept a place and get through the day. At Trantor I was on that team, on the backend under the guest flow and on the chatbot that answered when a person did not know where to go.

  1. Pass
  2. Ask
  3. Go
  • The chat covered the app and the sport: where the match is, when the tickets arrive, and the questions that show up the hour before kickoff.
  • Flights and hotels came back live through SerpAPI and Booking.com. A weather API and Google Places kept the answer tied to a ground, a forecast, and a place that was open.

BackendChatbotSerpAPIBooking.comWeatherPlaces

Visa GoMatch day

Where is the match, and when do I get the tickets?

Kickoff is at the ground on your pass. Tickets land tonight.

Find a flight and a hotel. What is the weather doing?

Clear skies. A hotel twelve minutes out is still open. Flights are live.

Places Weather Booking.com SerpAPI

Path

Five companies.

2024 – now · Visa Go team

Trantor

Visa Go · FIFA World Cup 2026

Trantor puts enterprise AI into production. I was on the Visa Go team: the backend and the chatbot for an invite-only app Visa's guests used at the World Cup. Visa sponsored the tournament. The app is how the guest list actually moved.

  • Visa's sponsored tickets covered about 30,000 people. Visa Go is where a guest accepted the place, built the trip, and asked what to do next.
  • My part of the backend: tickets, schedules, and the path from "I have a ticket" to "where do I go."
  • My part of the chat: where the match is, when the tickets arrive, and the questions that show up the hour before kickoff.
  • Flights and hotels came back live through SerpAPI and Booking.com. A weather API and Google Places kept the answer tied to a ground, a forecast, and a place that was open.

BackendChatbotSerpAPIBooking.comWeatherPlaces

The scrape

National brands in. Our own model out.

These are the sites the spiders ran. Click one. Fuel, parking, homes, and jobs took the same path, and the clean rows were what the LLM trained on.

  1. Collect Scrapy on a schedule. Fuel, parking, homes, and jobs each had their own job.
  2. Clean Raw pages landed in MongoDB. ETL typed the prices, places, hours, and listing text.
  3. Train Those tables trained our own LLMs. RAG so an answer came back from our rows.

Shell · fuel

Station pages and pump prices, on a schedule. Raw pages in MongoDB, then ETL into a typed price and a place. Those rows trained the LLM, and RAG retrieved them later.

Contact

Tell me what you are building.

The note opens in your email app, addressed to me. Nothing is stored on this site.