AWS Builder Center

Building CloudVault AI — What I Learned Combining S3, Lambda, and Bedrock in a Real Project

Built CloudVault AI, a cloud-native doc/media platform with AI features — like Google Drive but with auto-summarization, AI chat, and auto-generated study material (MCQs/flashcards). Stack: S3 + Lambda (Python) branching by file type, Textract for PDFs, Bedrock (Nova Lite) for images/AI, DynamoDB, Cognito auth, API Gateway with JWT authorizer. Key lesson: switched from Claude to Nova Lite after hitting an AWS Marketplace payment requirement — Nova bills natively, no extra setup needed.

Yep — just copy the whole body text below and paste it directly into the article body field:


Building CloudVault AI — What I Learned Combining S3, Lambda, and Bedrock in a Real Project
As a final-year Cloud Computing student, I wanted to go beyond tutorials and build something end-to-end — so I created CloudVault AI, a cloud-native document and media platform that works like Google Drive but with AI baked in: automatic summarization, per-document AI chat, and auto-generated study material (MCQs + flashcards) for any file you upload.
The Stack
  • Amazon S3 for storage, with per-user key prefixing for isolation
  • AWS Lambda (Python) functions triggered by S3 uploads, branching by file type — plain text read directly, PDFs processed via Amazon Textract, images analyzed via Amazon Bedrock (Nova Lite)
  • Amazon DynamoDB for documents, folders, and shareable links
  • Amazon Cognito for authentication with a fully custom sign-up/sign-in UI
  • API Gateway (HTTP API) with a JWT authorizer fronting everything
Real Lessons Tutorials Don't Usually Mention
  1. S3 event notifications URL-encode filenames — spaces become +. You need to decode them (urllib.parse.unquote_plus) before using them in Lambda, or file lookups silently fail.
  1. Presigned upload URLs are strict. The ContentType used when generating the URL must exactly match the browser's PUT request header, or S3 rejects the upload with SignatureDoesNotMatch.
  1. New Lambda functions don't inherit your shared IAM role by default — they get their own auto-generated execution role. Easy to miss until you hit a confusing AccessDeniedException on a function that "should" have permissions.
  1. Model switching due to billing setup: I originally wanted to use Claude via Bedrock for the AI features, but hit an AWS Marketplace payment-instrument requirement for Anthropic models. Switching to Amazon Nova Lite solved it instantly, since it bills natively through AWS with no additional setup required.
  1. CORS can fail silently in two places at once. I had CORS headers set in both Lambda and API Gateway, which caused duplicate headers and browser rejection — even though the backend response itself was successful.
What's Next
  • Async Textract support for multi-page PDFs (the sync API only handles single-page documents)
  • AI processing for video files, which are currently stored but not yet analyzed
Building this taught me more about real-world AWS service integration than any single course could — especially the small, undocumented gotchas that only show up once you're wiring services together yourself.
Happy to go deeper into any part of the architecture — feel free to comment if you're building something similar or hit similar issues!
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