for AWS Community Builders

Posted on Oct 25

Building Gulf Heritage AI Studio: A Serverless GenAI Experience

#aws #genai #ai #serverless

How I leveraged AWS SAM, Lambda, Bedrock Nova Canvas, Rekognition, and Translate to create an AI-powered cultural heritage poster generator.

Introduction

For GITEX 2025, the goal was to build an AI-powered experience that transformed attendee selfies into heritage-themed posters, each with rich UAE or KSA cultural motifs and personalized Arabic calligraphy. The project became a deep dive into scalable serverless architecture, generative AI services, and user privacy—all executed within a few short weeks.

The Challenge

Process selfies in real-time
Generate authentic, high-quality backgrounds
Auto-scale and require minimal operational overhead
Enable easy poster sharing and instant download

A serverless stack using AWS and client-side image segmentation enabled meeting every technical requirement and business need.

Architecture: Modern AWS Serverless Stack

┌───────────────────────────────────────────────┐
│               CLIENT SIDE                    │
│ ┌───────────┐ ┌───────────────────┐          │
│ │  Webcam   │→│  MediaPipe (JS)   │          │
│ └───────────┘ └─────┬─────────────┘          │
│          (selfie capture/segmentation)       │
└──────────────┬───────────────────────────────┘
               │  Base64 (segmented cutout)
               ▼
   HTTPS POST to AWS Lambda Function URL
               │
┌──────────────▼───────────────┐
│           Lambda             │
│ ┌─────────┬───────────────┐  │
│ │Bedrock  │ Rekognition   │  │ (parallel)
│ │Nova     │ (contextual   │  │
│ │Canvas   │ design)       │  │
│ └─────────┴───────────────┘  │
│        │                     │
│        ▼                     │
│   AWS Translate              │
│   (Arabic name)              │
│        ↓                     │
│      Image compositing       │
│        ↓                     │
│      S3 Poster storage       │
│        ↓                     │
│    QR code Generation        │
└─────────┬──────────────────┘
          │
          ▼
 Client receives: Poster URL + QR Code

Under the Hood

Amazon Nova Canvas: Text-to-image + image variation for heritage background generation
Amazon Rekognition: Intelligent attribute detection (gender) for contextual design
Amazon Translate: Seamless English→Arabic name translation
AWS Lambda: Serverless image processing and compositing
Amazon S3: Secure storage with presigned URLs
MediaPipe: Client-side face segmentation (privacy-focused)

Implementation & Deep Dive

Client-Side Security

MediaPipe (JavaScript) segments selfies in-browser.
Privacy-First: Only cutouts (not full selfies) are sent to the backend.

Serverless Deployment

AWS SAM automates all infrastructure as code.
Lambda orchestrates: receives the base64 cutout, generates personalized backgrounds via Bedrock Nova Canvas, uses Rekognition only for context (not ID), and overlays Arabic names using AWS Translate, compositing everything into a high-resolution poster.

Key Lambda Pipeline (Python-like pseudocode):

1. Decode base64 cutout
2. Parallel: (a) Amazon Rekognition, (b) Amazon Nova Canvas generates background .
3. AWS Translate: convert English name to Arabic
4. Composite images, add branding, stylize name
5. Store poster in S3, generate presigned download URL
6. Generate + store QR code linked to poster
7. Return both URLs to client for download/sharing

Code Samples

Amazon Rekognition Code :

def detect_gender(image_bytes):
    try:
        response = rekognition.detect_faces(Image={'Bytes': image_bytes}, Attributes=['GENDER'])
        if response['FaceDetails']:
            face = response['FaceDetails'][0]
            gender = face['Gender']['Value'].lower()
            confidence = face['Gender']['Confidence']
            if confidence > 80:
                return gender
            else:
                return 'female'
        else:
            return 'female'
    except Exception:
        return 'female'

Amazon Translate code :

def translate_to_arabic(english_name):
    try:
        response = translate.translate_text(Text=english_name, SourceLanguageCode='en', TargetLanguageCode='ar')
        return response['TranslatedText']
    except Exception:
        return english_name

Thematic Authenticity

UAE Themes

Classic: Dubai skyline, Sadu patterns
Modern: Dubai Marina, contemporary cityscape
Desert: Golden dunes, Bedouin details

Each theme prompt was carefully engineered for Bedrock to ensure cultural accuracy and visual excellence.

Performance and Cost

Median end-to-end processing time: 8–12 seconds
Serverless scaling: no downtime or manual scaling
Total event cost: Effective cost for all compute, storage, and AI resources

Takeaways and What’s Next

Serverless is ideal: for bursty, unpredictable event traffic.
Bedrock Nova Canvas: delivers fast, high-quality custom heritage scenes with prompt tuning.
Future ideas: event-driven analytics, more languages, or NFT/AR integration.

DEV Community