By exploring smaller, more powerful AI systems that are explainable, transparent, and an asset to humanity - not a replacement.
Four properties, one intelligent system. Every project builds towards it.
Small in scale and powerful by design.
Trained in one domain and still usable in another.
Persistent in memory and continuous in learning.
Clear in reasoning and explainable by design.
A growing set of complementary open-source tools that stack toward one mission-driving framework.
A Flax-based Liquid RL research framework for small and powerful AI models. Combines Closed-form Continuous-time (CfC) Liquid Neural Networks with DiscoRL meta-learning and custom techniques to build adaptive AI agents.
An analytics platform for Velora agents. Provides live health evaluation, population monitoring, and inline metric descriptors across training runs. Supports DiscoRL meta-training and standalone CfC agent workflows.
A lightweight, Gymnasium-style API standard for JAX-native RL environments. Includes base classes, spaces, wrappers, and a shared registry. GPU/TPU-native, vmap-friendly, and fully jit-compilable.
A JAX-native MuJoCo environment suite for Envrax. Registers 25 DM Control Playground environments into Envrax's registry. Uses MJX-backed physics suitable for continuous control tasks.