Case study
AI content moderation architecture
A hybrid Trust & Safety design for a community platform: fast checks before content is published, deeper analysis after.
- Client
- TUVU
- Period
- 2025 – Present
- Role
- Architect and technology lead
Context
TUVU hosts user-generated posts, messages and images across 30+ communities, from universities and churches to creators and collectives. As the platform expanded internationally, content moderation became both a product requirement and a compliance one, with frameworks similar to the EU Digital Services Act in scope.
AI was the original reason I joined the company. From 2025, this is where it entered the platform.
Challenge
Moderation has to do two things that pull in different directions. Checks that run before something is published must be fast enough not to get in the user’s way. Deeper analysis is valuable but does not belong in the request path. A single mechanism could not serve both.
My role
I designed the architecture and led its implementation with the engineering team.
What I did
I split moderation into two layers.
Synchronous · before publication · Cloudflare Workers
- User submits post
- Text → OpenAI · Images → third-party service
- Published or blocked
Asynchronous · after publication · AWS Lambda
- Content stored
- AWS Rekognition + in-house models
- Flagged for action
Synchronous, before publication (edge)
- Text and messages: checked with OpenAI before they are posted.
- Images: checked with fast third-party services.
- Deployed on Cloudflare Workers, close to the user.
Asynchronous, after publication (cloud backend)
- A pipeline running AWS Rekognition on Lambda.
- Complemented by models trained in-house.
AI in how we build
AI also changed how the team works. I connected GitHub with Cursor for automated pull-request review. We use Claude as a coding assistant, and we now generate QA test cases with Claude.
Results
- AI-based moderation in production since 2025, covering text, messages and images.
- A clear separation between real-time checks and deeper background analysis, each on the infrastructure that suits it (edge workers and serverless functions).
- Moderation that supports the platform’s compliance reviews for international launches.
- AI-assisted development adopted across the engineering team, part of what let the team right-size from ~20 to 14 while increasing velocity.
Technologies
- OpenAI
- Cloudflare Workers
- Third-party image moderation services
- AWS Rekognition
- AWS Lambda
- In-house trained models
- Cursor
- Claude