More mathematics.
More possibility.
Derivon builds omnibias, an open-core foundation for derivatives and trainable decisions. We want scientific models to use the mathematics their problems demand.
Construct the computation.
Then test the claim.
Activation identities, Taylor composition, and direct contractions offer alternative paths through derivative-heavy problems. Our work connects those mathematical structures to trainable software.
We publish the implementation and the evidence together. A useful result includes its assumptions, numerical limitations, and comparison conditions.
Meet the project ↗Open at the foundation
Public code and package-specific licenses make the substrate inspectable. You can see exactly what runs inside your model.
Precise about the result
Closed-form, automatic differentiation, numerical approximation, and checked certificates carry distinct meanings.
Built for the training loop
Forward spatial derivatives retain parameter gradients on supported paths. The derivative becomes part of the objective.
Measured in context
Benchmarks identify the hardware, dtype, workload, and baseline. Accuracy is evaluated independently from speed.
Bring a difficult problem.
For integration work, research collaboration, commercial licensing, or interest in contributing, reach the team at info@derivon.ai.