Making high-order derivatives a cheap primitive
Derivon started as doctoral research on a simple question: why pay for nested automatic differentiation when the derivative of an activation has a closed form at every order? The answer became a derivative engine for scientific machine learning.
Our mission
Give every scientific ML team derivative operators that are fast, bit-stable, and certifiable - without rewriting their models.
Our vision
A world where the numerics under a scientific result are reproducible by default and verifiable on demand.
Derive + on
A derivative, treated as a fundamental particle of computation.
Written ∂erivon, the leading glyph is ∂ - the partial-derivative operator - standing in for the D. It reads as the name to everyone, and as the mathematics to those who know the symbol.
deriv + on - the canonical reading
“-on” is physics’ suffix for a fundamental particle: electron, boson, gluon. Derivon treats the derivative as a first-class particle of computation - the exact, closed-form quantum every higher operator is built from.
de + rivon - a quieter resonance
Split the other way and you meet the Latin rivus, a stream - fitting, because the engine is about derivatives that flow: Taylor jets propagating through a network, quantities streaming through typed operators. A coincidence we were happy to find, not a claim we planned.
Why Derivon works the way it does
Exact where the math permits
We never blur closed-form, autodiff-exact, and numerical. Each result carries an honest label, so you always know what kind of 'exact' you are getting.
Trust is earned, not asserted
Speed without reproducibility is marketing. Every claim ships with a benchmark you can re-run and, where it matters, a certificate you can check.
Open core, vendor neutral
The math is open and auditable. No scheduler, host, or framework lock-in. The same primitive runs on PyTorch, JAX, and Keras 3, bit-for-bit.
Depth in one thing, done rigorously
We are not a general AI platform. We are specialists in the one primitive that dominates scientific ML inner loops - and we go deep enough to certify it.
- Founded by the researcher who built the derivative-tower primitive
- Open-core, so you can audit the math instead of trusting a vendor
- Reproducible benchmarks - re-run the CPU smoke tier on your laptop
- Honest scope: we tell you when autodiff is already the right tool
From thesis to product
- 2023-2025
Doctoral research
The activation derivative-tower primitive is developed and validated: the Riccati / Eulerian / Hermite recurrences behind closed-form σⁿ(z).
- 2026 Q1
Open-core foundation
The stable workspace - core, PyTorch, JAX, and the FermiNet bridge - is hardened with cross-backend bit-parity and a reproducible benchmark tier.
- 2026 Q2
Scientific extensions
Physics-informed and field-substrate packages reach beta; certified-numerics, verification, and discovery packages land as labelled alpha.
- 2026 H2
Derivon, the company
Design-partner engagements formalize: derivative-acceleration audits, integration projects, and certified-numerics reports for scientific teams.
The people behind the engine
A small, senior team of numerical analysts and ML engineers. Profiles are representative of the roles we are building out.
Dr. Amaya Rendón
Founder & Chief Scientist
Numerical analyst focused on differentiable scientific computing and validated numerics. Built Derivon's closed-form derivative engine out of doctoral work on activation derivative towers.
Kenji Watanabe
Member of Technical Staff
Works on the JAX and PyTorch backends and the FermiNet integration path. Previously optimized kernels for large-scale variational Monte Carlo.
Lena Fischer
Head of Developer Experience
Builds the SDK, documentation, and reproducibility tooling. Cares about benchmarks you can re-run on a laptop.
Prof. Daniel Okonkwo
Scientific Advisor
Works on validated numerics and computer-assisted proofs. Advises on the rigorous and formal registers.
Build the math layer for scientific AI
We hire numerical analysts, ML systems engineers, and developer-experience specialists who care about getting the bits exactly right. Remote-first across North America and the EU.
Put closed-form derivatives in your inner loop
Send us one derivative bottleneck. We will benchmark a closed-form, bit-stable replacement against your current autodiff path - on your problem sizes.
Or email info@derivon.ai