Building an AI-Powered Photo Gallery with Serverless and Event-Driven Magic!
- 1. Building an AI-Powered Photo Gallery with Serverless and Event-Driven Magic Srushith Repakula
- 2. Who am I? Head of Engineering, KonfHub AWS Serverless Hero Organiser of AWS User Group Tirupati Co-organiser of Serverless Group
- 3. 01. What Are We Trying to Build? 02. How Did We Build It - The Architecture? 03. Under the Hood: The AI Sorcery 04. Optimisation & Performance 05. Scalability & Privacy 06. The Download Challenge 07. Key Takeaways CONTENTS
- 4. “ What are we trying to build? An intelligent photo gallery that can: ● Automatically compress uploaded images for beer performance ● Detect and index faces using AI (Amazon Rekognition) ● Let users search by face – no manual tagging needed! ● Scale eortlessly without managing any servers
- 5. How did we build it?
- 6. Under the Hood: The AI Sorcery Amazon Rekognition
- 7. “ Amazon Rekognition is an AI service from AWS that makes it easy to analyze images and videos using deep learning - no ML expertise needed!
- 8. “ ● Label Detection - Identifies objects, scenes, and activities in an image. ● Image Properties - Analyzes brightness, sharpness, and contrast of images. ● Image Moderation - Flags inappropriate or unsafe content in images. ● Facial Analysis - Detects facial aributes like age, gender, and emotions. ● Face Comparison - Compares two faces and returns similarity confidence. ● Face Liveness - Determines if a face in an image is from a live person (anti-spoofing).
- 9. “ Amazon Rekognition - IndexFace A Rekognition API that detects faces in an image and stores facial features in a face collection for future matching and search
- 10. “ What’s a Face Collection? ● Think of it as a searchable database of faces ● You create a collection (like a named folder) ● Each face is stored with: ○ A unique FaceId ○ Optional ExternalImageId (e.g., user or image reference) ○ Bounding box & facial landmarks
- 11. “
- 12. What is happening under the hood? This allows us to: 1. Search across images by uploading another face 2. Retrieve matching photos instantly 3. Do all this with no manual tagging or sorting
- 13. An API that compares a new input face against faces stored in a collection and returns the most similar matches with confidence scores. Amazon Rekognition - SearchImageByFace
- 14. ● Define the collection to search within ● Can specify match threshold (e.g., 90% confidence) ● Response includes: ○ FaceId ○ ExternalImageId ○ Similarity score ○ Bounding box Amazon Rekognition - SearchImageByFace
- 16. On-the-fly Image Compression Uploaded images are compressed using a Lambda function to: ● Reduce image size by up to 80% ● Improve load speed on the gallery view ● Save on bandwidth and storage costs CloudFront for Global Caching Serve all user-facing images via Amazon CloudFront, which: ● Caches compressed images at edge locations ● Reduces latency for global users ● Adds resilience and scalability
- 17. Instead of mapping face IDs back to the image manually: ● We store the image URLas ExternalImageId when calling IndexFaces ● This gives us a direct reference to the original photo for any face match ● Avoids the need for a separate DB lookup or reverse mapping Smart usage of ExternalImageId
- 18. ● Using Arm64 as the Lambda instruction set for beer cost savings and performance - 10-15% faster compression times & 20% cost reduction ● Parallel processing of image compression + indexing via SQS → Lambda triggers ● Use S3 transfer acceleration for uploads from the dashboard. Helps with faster uploads, especially when the original images are larger in size ● Seing up S3 lifecycle policy - Automatically transition or delete old objects using S3 lifecycle rules: Other optimisations
- 20. Built entirely with Serverless services like S3, Lambda, and Rekognition — which scale automatically with no infra provisioning Using SQS decouples processing workloads (compression & indexing), ensuring resilience and elasticity during spikes Each image triggers its own independent processing — no shared bolenecks or thread locks. Delivers compressed images globally with low latency regardless of upload volume. Scalability Serverless by Design Asynchronous & Event-Driven Parallelism in Image Workflows CloudFront Edge Caching
- 21. Original images uploaded from the dashboard are via signed URLs for beer security Rekognition doesn’t store actual images - only facial metadata. You retain full control of the original image in S3. Maintain separate Rekognition face collections per event, per environment Each Lambda only has scoped-down permissions (ex, no Rekognition write access unless needed) Privacy Signed URLs for S3 Access No PII Stored Beyond Face Features Face Collection Isolation IAM Role Isolation
- 22. Delivering thousands of high-resolution gallery images isn’t just about downloading files. It’s about doing it eiciently, reliably, and serverlessly! The Download Challenge
- 23. Delivering thousands of high-resolution gallery images isn’t just about downloading files. It’s about doing it eiciently, reliably, and serverlessly! The problems? ● Users want to download thousands of high-res images ● Large file sizes = risk of Lambda timeouts and memory limits ● Single large ZIPs are slow, unreliable, and hard to manage ● Needs to be serverless, scalable, and user-friendly The Download Challenge
- 24. ● Step Functions orchestrate the entire flow ● Images batched dynamically (~2GB per batch) to avoid Lambda limits ● S3 Map state + parallel Lambda downloads each batch, zips, uploads ● Email sent to user with all ZIP URLs after processing ● No server to maintain - purely event-driven! Our Serverless Solution
- 25. Key Takeaways ● Faster development, lower costs, and beer performance — without sacrificing intelligence ● You can build production-grade smart systems with surprisingly few moving parts ● AI integration doesn’t have to be complex ● Serverless architecture allows you to scale eortlessly, with no infra management ● Event-driven design is powerful for decoupling workflows and reacting to changes in real time Why It Matters What We Learned
- 26. “ “Moments captured. Faces remembered. Powered by AI, delivered by Serverless! “
- 27. THANK YOU! Srushith Repakula Head of Engineering, KonfHub AWS Serverless Hero