nesa

Why we included this project

Teams that must run large language or image models over sensitive data usually have to choose between speed and privacy. Nesa sidesteps that tradeoff with its own equivariant encryption scheme, which keeps inference latency essentially unchanged and covers common activations and normalizations, so the hosting provider never sees the inputs or even the query. The service exposes a ChatGPT-compatible API, so a developer can switch with one line of code, and it supports models such as Llama, Mistral, and Stable Diffusion. For regulated or confidential data, that combination of near-zero overhead and end-to-end privacy is a genuinely unusual offer in the AI API space.