derivation.

Security Model training & eval

Specialized models
for cybersecurity.

We build security models and the environments
to train and evaluate them.

AI is scaling software production and attacks. We believe specialized models, trained with verifiable feedback, can improve security task performance while reducing inference costs.

Our work brings together task-specific post-training, reproducible environments, and evaluation against real-world outcomes.

Founding team

Research meets
engineering.

Our experience spans large-scale model training, reasoning, and automated verification.

Kaiyu Yang

Reasoning capabilities

Led reasoning capability development for a 1.5T-parameter MoE model at Apodex, spanning pre-training, mid-training, and post-training.

Princeton CS PhD. Former Caltech postdoc. Formerly Meta FAIR. LeanDojo first author.

Chenyang An

RL environments & evaluation

Developed RL environments and evaluation systems at Apodex. Previously built and productionized policy verification systems at AWS Automated Reasoning.

UC San Diego Mathematics PhD. Formerly AWS Automated Reasoning Group.

Zenan Li

Training infrastructure

Senior individual contributor on training infrastructure for a 1.5T-parameter MoE model at Apodex. Experienced in training and inference optimization.

Nanjing University CS PhD. Former ETH Zurich postdoc. Goedel-Code-Prover co-first author.