AWS Builder Center
BankMate: A Voice-First Banking Kiosk for People Who Can't Read or Write

BankMate: A Voice-First Banking Kiosk for People Who Can't Read or Write

How I built a camera-and-voice-driven kiosk prototype on AWS that lets anyone withdraw cash or update their mobile number without reading or writing a single form.

The moment this idea started

A few months ago, I went to the bank with my grandmother. While we waited, I noticed something I'd never really paid attention to before a row of elderly people sitting there, not because they couldn't bank, but because they were waiting for someone to help them fill out a form. They couldn't read or write, and every withdrawal or update meant depending on a busy staff member to do it for them.
That stayed with me. So I built BankMate.

What BankMate does

BankMate is a kiosk prototype think of it like an ATM, but for paperwork instead of cash. It's meant to sit inside a real bank branch, not be downloaded as an app. The flow is simple:
  1. Scan your passbook with the kiosk's camera
  1. Speak your request in Hindi or English
  1. The kiosk understands and auto-fills the correct bank form
  1. You sign ,that's the only thing left for you to write
It currently supports two tasks: cash withdrawal and mobile number updates. Every screen also has a one tap "Call Staff" option BankMate isn't trying to replace bank employees, just make reaching one easier when a human is genuinely needed.

The AWS architecture

Everything runs on real, deployed AWS infrastructure in Mumbai (ap-south-1), entirely within the Free Tier:
  • Amazon S3 temporarily holds the passbook photo during processing, then deletes it immediately after nothing about the customer sits on a server afterward.
  • AWS Lambda (Python) handles the upload, extraction, and cleanup logic, sitting behind Amazon API Gateway as a clean HTTP API.
  • AWS Amplify Hosting serves the React frontend over HTTPS which turned out to be more than a hosting convenience. Camera and microphone access in the browser require a secure HTTPS context, so Amplify is what actually makes the app usable outside local development.
  • IAM scopes a dedicated Lambda role to only the permissions it needs S3, Textract, and Bedrock.
The architecture was designed around Amazon Textract for reading the passbook and Amazon Bedrock's Nova Micro for understanding voice commands. Both Lambda functions are already wired to call them directly.

The honest part: hitting a wall with a new account

Textract and Bedrock are currently gated on my AWS account during new-account identity verification confirmed on two separate accounts, so it's a platform-level restriction, not a bug in my code. I decided not to spend the rest of my limited hackathon time fighting this, and instead built around it:
  • Passbook extraction currently uses realistic mock data
  • Voice understanding runs on the browser's own speech recognition, with keyword matching across English, Hindi, and Hinglish
The key decision I made here: both substitutes return data in the exact same shape the real AWS services would. That means swapping in Textract and Bedrock later needs zero frontend changes only the Lambda internals change once verification clears.

A deployment bug worth sharing

One bug ate more of my time than I expected, and it's worth mentioning since it's an easy trap for Windows-based builders using Amplify's drag-and-drop deploy.
After deploying my built React app, every single asset CSS, JS, even images 404'd, despite the build being completely correct. It turned out Windows' built-in zip tools write internal zip file paths using backslashes , which is the normal Windows path separator. Amplify's hosting runs on Linux, which expects forward slashes as folder separators so it read  assets\index.css  as one literal filename with a backslash character in it, not as a file inside a folder.
The fix was to build the zip manually in PowerShell using .NET's  ZipArchive  class, explicitly normalizing every path to forward slashes before adding it to the archive. A small thing, but it cost real debugging time before I traced it back to the zip's internal structure.

What's next

  • Swap in real Textract and Bedrock the moment account verification clears
  • Add regional languages beyond Hindi and English as it scales
  • Explore physical kiosk hardware a proper mic array, thermal printer, tamper-resistant enclosure for actual branch deployment

Closing thought

BankMate started from watching my grandmother wait at a bank branch. It's still a prototype, but it's built on real, working AWS infrastructure, with an honest account of what's live today and what's designed to plug in the moment access allows it.


Built for the Bharat Builds Tour: First Commit hackathon (wemakedevs.org) .

Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article