What it is
Neutryx introduces a general-purpose constrained computation engine: a neural system that learns to approximate arbitrary functional relationships under explicit mathematical, physical, or probabilistic constraints. Rather than hard-coding solvers, constraints are embedded and enforced through a bilevel adaptive Lagrangian optimiser with meta-gradient control.
Why it matters
- Domain-agnostic: physics, finance, engineering — one engine.
- Model-free: no predefined operators or meshes.
- Admissible by design: conservation, no-arbitrage, positivity and normalisation are enforced as constraints.
- HPC-ready: JAX with JIT, vectorisation and pmap/pjit; GPU/TPU acceleration.
Dual-Layer Release Model
We separate openness for research from protection for commercialisation:
- Neutryx Core (Public): open-source JAX library with differentiable pricing, PDE solvers, AAD/JIT, Monte Carlo, documentation and demos.
- Neutryx AI Engine (Proprietary): constraint-coupled bilevel learning, dynamic constraint embeddings and meta-gradient controllers — the subject of our UCLB invention disclosure.
The Core demonstrates reproducibility and scientific validity; the Engine contains the protectable IP.