Case study
Building Aether AI Support from scratch.
How we built a modern, AI-powered support platform from the ground up — cutting-edge technology on a cloud-native architecture.
The challenge
Support software was broken.
- Slow response times leading to customer frustration and churn
- Repetitive queries consuming valuable agent time
- Lack of intelligent routing causing mishandled tickets
- Disconnected systems creating data silos
- Complex pricing making it inaccessible for SMBs
We set out to fix all of it with AI — while staying affordable and dead-simple to deploy.
Our solution
Four decisions that defined it.
AI-first architecture
A Retrieval-Augmented Generation pipeline on GPT-4o-mini with pgvector embeddings. The AI understands context, retrieves the right knowledge, and generates accurate answers in real time.
Real-time collaboration
WebSockets and Server-Sent Events deliver instant updates platform-wide — live ticket changes, typing indicators and presence, with no refresh.
Multi-tenant SaaS design
Unlimited companies with complete data isolation. Each tenant customizes widget, knowledge base, SLA rules and integrations independently.
Serverless infrastructure
Backend on Google Cloud Run, frontend on Vercel — automatic scaling, negligible cold starts, and 99.95% uptime with no servers to manage.
Technical deep dive
Under the hood.
Backend architecture
FastAPI for its exceptional performance (40,000+ req/s) and native async. Fully typed with Pydantic for automatic validation and OpenAPI docs. • Service-oriented design with dependency injection • Background task processing with asyncio • Structured logging with correlation IDs • Rate limiting and quota enforcement
AI & machine learning
The RAG pipeline processes each query in stages: 1. Query analysis — intent detection and entity extraction 2. Retrieval — hybrid semantic + keyword search 3. Context assembly — relevance scoring and window optimization 4. Generation — GPT-4o-mini with per-company system prompts 5. Post-processing — confidence scoring and escalation detection Function calling lets the AI escalate to a human seamlessly when needed.
Frontend engineering
Next.js 14 App Router for SSR, streaming and optimal bundle splitting: • Context-based theming system • Framer Motion for smooth UX • Ctrl+K command palette • Optimistic UI with SWR caching • Responsive Tailwind layouts
Database design
Neon PostgreSQL — serverless relational storage with automatic scaling: • Companies — multi-tenant configuration • Tickets — support requests with full history • Messages — conversation threads with metadata • Knowledge base — articles, FAQs, pgvector embeddings • Users — agents, admins, customers Indexed SQL for complex lookups, pgvector for semantic search.
Results
The numbers that matter.
Lessons learned
What we'd tell our past selves.
Start with the AI, not the UI
We built the RAG pipeline first, then designed the interface around it — so the AI was a core feature, not an afterthought.
Invest in real-time from day one
Retrofitting WebSockets is painful. Designing for real-time upfront avoided technical debt and delivered a more responsive product.
Multi-tenancy requires early planning
Every query, route and component needed tenant awareness. Planning it upfront saved months of refactoring.
Type everything
TypeScript and Pydantic caught countless bugs before production. Strong typing is non-negotiable for complex apps.
KumoKodo.ai
We build AI-powered SaaS on modern cloud-native technology — combining machine learning, full-stack engineering and UX design to ship products that solve real business problems.
Want to try Aether?
Start a 14-day free trial and feel AI-powered support for yourself.