Building an AI Expert Site for Under $25/Month
Deploying a production-ready AI vertical site doesn’t require an enterprise budget or thousand-dollar monthly cloud spending commitments. The Expert RAG architecture was deliberately designed to minimize baseline infrastructure costs while providing enterprise-grade streaming performance.
Here is the exact cost breakdown for running a live domain expert site at test-to-launch scale.
Infrastructure Cost Breakdown
| Component | Provider | Tier / Configuration | Monthly Cost |
|---|---|---|---|
| Frontend UI & Edge | Cloudflare Pages | Global Static CDN + Functions | $0.00 |
| Backend & Ingestion Worker | Railway | Web Service + Async Worker | ~$10.00 - $15.00 |
| Database & Vector Store | Supabase | Free Tier / Postgres + pgvector | $0.00 |
| Object Storage | Railway Bucket | Document Storage | ~$0.50 |
| LLM & Embedding Usage | OpenRouter / OpenAI | Pay-per-token API usage | ~$10.00 - $20.00 |
| Custom Domain | Registrar | .com / .org domain | ~$1.00 |
| Total Baseline | ~$21.50 - $36.50/mo |
How We Keep Baseline Costs Under $25
1. Static-First Edge Deployment
The customer-facing application UI runs as static HTML/JS on Cloudflare Pages. Cloudflare offers unlimited requests and bandwidth on its free tier. There are no server instances idling 24/7 just to serve static assets.
2. Browser-Direct Document Uploads
When an operator uploads a 50MB medical PDF, the document does not flow through the API server. The browser requests a presigned upload URL and streams the file directly to object storage. This reduces compute overhead on the backend web server, allowing it to run on minimal resource allocations.
3. Unified LiteLLM Gateway
All model calls pass through LiteLLM. Changing default models—for instance, switching between Anthropic Claude Sonnet, OpenAI GPT-4o-mini, or Llama 3 via Groq—requires updating a single line of configuration in vertical.toml. Operators can choose cost-efficient models for low-complexity domain queries.
4. Zero Lock-In Demolition
Because each vertical is packaged as an isolated deployment stack, operators can pause or tear down a non-performing domain site with a single click. There are no shared multi-tenant database migration headaches or recurring platform maintenance fees.
Infrastructure costs scale predictably with user activity. At launch scale, an operator can validate domain-market fit for under $25 per month.