The AI/ML stack we ship to clients in 2026

For 90% of client AI/ML work we ship: Python 3.11 + FastAPI for the model API, scikit-learn or PyTorch depending on the problem, Hugging Face for any transformer-based stuff, Postgres + pgvector when we need retrieval, and a tiny React UI for the demo.

Why this combo? FastAPI gives us OpenAPI docs for free. Hugging Face means we don't reinvent transformers. pgvector keeps everything in one database. And React lets non-technical clients actually click around and feel the value. The whole thing deploys to a single VPS for under ₹600/mo.