AWS Content Moderation System
Real-time video content moderation platform powered by AWS Rekognition, with serverless architecture (CDK, Lambda, Kinesis, DynamoDB, CloudWatch). Processes video at 30 FPS with under 200ms latency and 50+ content categories.
My Role
Backend & Infrastructure Engineer
Core Stack
AWS CDK, Lambda, Rekognition, Flask, Python, Node.js
Timeline
2024.11 — 2025.01
Live Project
GitHubThe Challenge
The core problem
Organizations need to moderate video content in real time—catching inappropriate material before it reaches users. Human review is slow, expensive, and doesn’t scale for live streams.
Product vision
A serverless, automated system that processes video frames at scale with low latency and manageable cost—so moderation happens in real time without a human in the loop.
What I owned
Architecture and implementation: AWS Rekognition for frame analysis, serverless pipeline (Lambda, step functions or similar), frame-sampling strategy to hit 30 FPS and sub-200ms latency while controlling cost. Designed for scale and reliability.
Results
Real-time video moderation without human reviewers. Sub-200ms latency; scalable serverless design; lower cost than manual review.
System Architecture
Serverless, event-driven pipeline: video ingestion (Kinesis), ML moderation (Rekognition), storage (DynamoDB), and monitoring (CloudWatch).
Frontend
- ES6+ JavaScript with event-driven updates (CustomEvent API).
- Chart.js for confidence metrics and moderation trends.
- Responsive dashboard with live video feed and results panel.
- REST integration with Flask backend; 1s refresh for real-time metrics.
Backend & API
- Flask app for video capture (OpenCV) and API endpoints.
- AWS SDK (boto3): Rekognition, DynamoDB, CloudWatch, SNS.
- Lambda (Node.js/TypeScript) for Rekognition and stream management.
- API Gateway for stream create/delete; EventBridge for scheduled processing.
AWS Services
- Amazon Rekognition: 50+ moderation labels, confidence scores.
- Kinesis Video Streams, DynamoDB (GSI on timestamp), CloudWatch, SNS.
- AWS CDK (TypeScript) for IaC; multi-stack (KVS/Rekognition + API/Lambda).
Real-Time Moderation Dashboard
Live video feed with frame-by-frame processing display, real-time moderation results with color-coded confidence levels, and Chart.js visualizations for confidence trends. Alert notifications for flagged content with detailed labels.
Serverless Pipeline & Stream Management
AWS CDK Infrastructure as Code with multi-stack architecture. EventBridge-triggered Lambda for stream processing; Kinesis Video Streams for ingestion. Create/delete streams via API Gateway with automatic cleanup of EventBridge rules.
Project Impact
Real-time content analysis with high accuracy and low latency; serverless design for scale and cost efficiency.
<200ms
Latency per frame
99.9%
Uptime (serverless)
80%
Infra setup time reduced (CDK)
50+
Content categories detected
Reflection
I’d add configurable sensitivity thresholds and retention policies for flagged content, plus more granular cost dashboards per stream. Designing the Kinesis–Lambda–Rekognition pipeline with CDK was very instructive for serverless ML systems.