Case study · our own live product
Round Table: a production multi‑agent AI platform, built end to end.
Round Table puts Claude, ChatGPT, Grok and Gemini into one conversation. We designed and built every layer, from the product to the AWS infrastructure, and it runs in production at round-table.ai.
What it does
Ask once. Four models answer in one thread.
Comparing AI models usually means pasting the same question into four browser tabs. Round Table asks them all in parallel, and you can @mention one model when you only want one.
Architecture
Every layer, on AWS, deployed from code.
- Front end
- React, Vite, TypeScript, Tailwind CSS and Zustand.
- API
- Fastify on Node.js and TypeScript, in containers on ECS Fargate.
- Sign-in
- AWS Cognito.
- Real-time
- Socket.io over WebSockets, streaming each model as it answers.
- Data and files
- Aurora PostgreSQL with connection pooling; S3 for image attachments.
- Delivery
- AWS CDK for the infrastructure, GitHub Actions for CI/CD, CloudFront for static assets.
Engineering decisions
Three decisions worth talking about.
Streaming several models at once.
Each provider answers at its own pace, so every answer streams to the browser as it arrives, side by side in one thread.
Keeping long conversations usable.
Context-window tracking with automatic compaction keeps long threads within each model's limits.
Usage tracking built in.
Token accounting is part of the platform, not an add-on, so AI usage and cost stay visible.
What it proves
The same pieces your project needs.
A customer portal, plant-floor dashboard or internal AI tool sits on the same foundations.
Related:AI integration ·Web applications ·AWS cloud & DevOps
Want the architecture walk‑through?
On a 30-minute scoping call we'll show how Round Table is built and map it to your systems and your project.