deschain@research:~$ whoami
We build tools that make trained machine-learning models understandable, verifiable and repairable.
A trained neural network is a program nobody wrote down. We study how these models compute their answers internally, and turn what we find into exact, testable descriptions: specifications a person can read and a machine can check.
The work sits where three fields meet: machine-learning interpretability, formal verification and compiler engineering.
Models are deployed long before anyone can say why they behave the way they do. Teams who build on them need to know what a model inherited, why it fails when it fails, and whether a fix really works before they ship it.
An early-stage, founder-led research company. We work with publicly released, openly licensed models, run our experiments on cloud GPUs, and publish once results have been reviewed.
Founded by Philipp Deschain, a game developer and systems engineer. Games taught us that a complex system can be trusted only when its rules are explicit and its behaviour can be replayed exactly. We hold machine-learning models to the same standard.
philipp@deschainresearch.com
Leave your details and we will write when there is something to share.