AWS Builder Center

My Dadi Can't Use Shopping Apps. So I Built One With AWS That Talks Back. Team Suno — First Commit, WeMakeDevs × AWS Bharat Builds Tour

A voice-first shopping and warranty assistant for Hindi and English speakers, built on AWS Lambda, API Gateway, S3, DynamoDB, and Textract during the First Commit hackathon.

Student
TL;DR: SunoCart is a voice-first shopping and warranty assistant, built solo over one hackathon, running on AWS Lambda, API Gateway, S3, DynamoDB, and Textract. This post covers what it does, the architecture, and four real bugs I hit deploying it — including one AWS account restriction I had to build around less than a day before the deadline.

The problem I actually wanted to solve

My dadi wants to buy a mixer grinder. She knows exactly what she needs — "something good for the house, under ₹2,000" — and saying that out loud is easy. Typing it into a shopping app, reading through spec sheets, and comparing options is not. Months later, when something breaks, the invoice is buried in a message thread somewhere and nobody remembers if it's still under warranty.
That's the actual problem SunoCart tries to solve — not "let's add AI to shopping," but "let's make shopping, and remembering what you bought, less dependent on being comfortable with apps."

Why not just voice search?

Worth addressing directly: Amazon's own app already has Hindi voice search. I'm not pretending otherwise. What it doesn't do is understand a budget constraint as a real filter, explain a product back to you in plain language, or remember anything about what you bought once you've paid — especially if you didn't buy it on Amazon.
So SunoCart isn't a voice-search demo. It's two connected pieces:
  • VoiceCart — speak a request in Hindi or English, get filtered results and a spoken, comparative explanation, not just a results grid.

  • WarrantyWallet — upload a photo of any invoice, from any seller, and the app extracts the product, purchase date, and warranty details, lets you confirm or correct them, and saves them so you can ask about them later, by voice, months from now.

The connective tissue is what makes it one product instead of two features: discover → buy → save the invoice → ask about it whenever you need to.

The AWS architecture

  • Amazon S3 — stores invoice images privately, and also hosts the frontend
  • Amazon Textract — extracts raw text from uploaded invoice photos
  • Amazon DynamoDB — stores confirmed warranty records
  • AWS Lambda — runs the entire Flask backend, wrapped with a lightweight WSGI adapter so the application code didn't need to be rewritten for a serverless environment — it also ended up serving the frontend itself, more on why below
  • Amazon API Gateway — routes everything to that Lambda function over HTTPS
One honest note: Amazon Bedrock access was blocked at the AWS account level a few days before submission. The organizers confirmed Bedrock isn't required for this track, so the intent-extraction and invoice-structuring layers run on rule-based logic instead — built with the exact same input/output shape a Bedrock call would have, so it's a one-line swap whenever access clears.

What fought back

Voice input in Hindi returned English results. Real speech recognition transcribes Hindi into Devanagari script (मुझे, हेडफोन), not the Romanized text I'd first tested with (mujhe, headphone). My language detection only recognized the Romanized form. Fixed by detecting Devanagari Unicode ranges directly, and mapping Devanagari product words to the same catalog categories as their English equivalents.
Packaging Python dependencies for Lambda, from Windows. Lambda runs Linux; some Python packages compile differently per OS. Rather than install dependencies normally, which would grab Windows binaries that silently fail on Lambda, I force-downloaded Linux-compatible wheels directly with pip's  --platform  flag — no Docker, no separate machine needed.
API Gateway's default event format didn't match my WSGI adapter. Fixed by explicitly setting the integration to payload format 1.0 — found by testing the adapter against a simulated event before deploying anything, which caught the mismatch early instead of live.
CloudFront required account verification that wasn't going to clear in time. With the deadline close, I couldn't wait days for that. Instead of CloudFront in front of S3, I had the same Lambda function serve the frontend's HTML/CSS/JS directly through Flask — one deployed service instead of three, which sidestepped the block entirely and also removed any CORS complexity, since everything now runs from a single origin.

What's next

Given more time, this MVP uses a single shared demo account — per-user accounts via Amazon Cognito would be the natural next addition. Beyond that: swap the rule-based layer for real Bedrock once access clears, support more Indian languages, and let WarrantyWallet handle multi-page and PDF invoices through Textract's async APIs.

Try it

Built solo, over one very long hackathon, by Team Suno. #BharatBuilds
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