All three run on a free Hugging Face Space — if it has gone idle, the first load takes a few seconds to wake up.

4 areas, 50 tools

Architecture

One request, end to end

What happens between clicking “Try it” on a tool and getting a result back. Every count here is read from the codebase at build time, not written by hand.

1. Client01

   Browser

   Next.js · Vercel

   • 50 tool pages + 3 platforms
   • Some tools run fully in-browser
   • MediaPipe · ONNX Runtime Web
2. Gate02

   Safeguards

   6 modules, one concern each

   • Origin allowlist (hard 403)
   • Rate limit — 60/min, 10/min LLM
   • 10 MB request body cap
   • YARA scan on uploads
   • Daily spend cap · event log
3. Server03

   FastAPI

   ml-api · Hugging Face Space

   • 55 routers
   • Sibling services: SQL, EDA, Vision
   • Docker, CPU-only
4. Compute04

   Models & Stores

   Local first, API when needed

   • ONNX Runtime · PyTorch · timm
   • scikit-learn · XGBoost · LightGBM
   • spaCy · sentence-transformers
   • ChromaDB vector store
   • 4 LLM providers, budget-capped

Requests flow left to right; results return the same way. The safeguard layer is the only thing between the public internet and the API, so it runs on every request regardless of which tool made it.