A Constraint-Driven, LLM-Augmented Framework for Multi-Objective Cloud Architecture Synthesis and Cost Optimization
Describe your app in plain English → get 2–3 cost-optimal cloud architectures, with real pricing, the money-saving optimization techniques most people don't know about, a diagram, and deployable Terraform.
Plain English
→ LLM extracts Requirement JSON
→ Engine + Knowledge Base picks services & optimizations
→ Generate Terraform (from vetted modules)
→ Infracost prices it ← REAL numbers, not LLM guesses
→ mingrammer/diagrams renders it
→ 3 options: Cheapest / Balanced / Most Reliable + cross-cloud comparison
| Mode | Entry point | Output |
|---|---|---|
| Design | "I'm building something new" | 2–3 priced architectures + diagram + Terraform |
| Advise | "I have a question" | Grounded answers citing real pricing + techniques |
| Audit | "I'm overpaying" | Upload billing CSV → waste report with $ saved/month |
Existing tools either draw pretty diagrams (no cost) or calculate prices (no reasoning). Competitors' cost numbers come from an LLM guessing. Ours are computed by a real pricing engine.
The moat is knowledge-base/ — the curated catalog of optimization techniques
(Graviton, zram, spot+checkpointing, scale-to-zero, egress-free storage…),
each mapped to when it applies, what it saves, and the tool that implements it.
WhichCloud/
├── docs/ # PRD, resource map
├── knowledge-base/ # ⭐ THE MOAT — curated optimization techniques
│ ├── techniques/ # one YAML per technique
│ └── service-mappings/ # AWS ↔ GCP ↔ Azure equivalents
├── backend/ # FastAPI — engine, RAG, cost, generators
├── frontend/ # Next.js — chat, results, diagrams
└── infra/ # docker-compose: Postgres, Redis, cloud-pricing-api
🚧 Pre-v1. Scaffolding stage. Next step: verify the cost pipeline
(self-host cloud-pricing-api, run infracost breakdown on sample Terraform with no cloud credentials).