Vatana AI Platform
An Apple-Silicon-first research platform of measured engines, one public index
2026Owner, researcher, and builder

01 / Overview
Overview
Research platforms drift when the best model is hard-coded and fallbacks stay hidden, so this one is a set of specialized engines that each have to earn their place: a challenger becomes the champion only by beating the reference on measured evaluation. It runs Apple-Silicon-first on PyTorch and MLX with experiments tracked in MLflow, and the repository is the source of truth; the public page is a thin index projected from that state and held in sync by a check.
02 / Stack
Stack
Python 3.13
uv
PyTorch
- Apple MLX
MLflow
Next.js
Vercel
03 / Where it stands
Where it stands
The public engine index is live on Vercel; the platform itself runs on the owner's machine and publishes no model, dataset, or hardware detail beyond that page. Its gateway and a CPU route for the Khmer champion are built and measured (CER 0.165 on CPU against the champion's 0.148, same 150 utterances) and not yet deployed.
Working on something similar?
Tell me what you're building.