I build systems that autonomously find failure modes in Large Language Models using Reinforcement Learning.

Currently, I am working at KachmanLab on an automated jailbreaking framework. My work focuses on the intersection of theoretical alignment research and high-performance engineering.

• Reinforcement Learning: Implementing verifiable reward frameworks (using GRPO and Verifiers) to train adversarial agents.
• Inference Optimization: Designing asynchronous generation pipelines using asyncio and vLLM to maximize throughput across multi-GPU environments.
• Evaluation: Curating adversarial datasets and benchmarking model robustness using AgentDojo and custom environments.

• KachmanLab (Current): End-to-end RL training for automated jailbreaking.
• Prime Intellect: Developed RL environments for decentralized training (e.g., Gutenberg literary analysis).
• RoboLearn: Built Bayesian models for depression (controllability) and data analysis pipelines for computational psychiatry in collaboration with Prof. Dr. Roshan Cools (Donders Institute/RadboudUMC).

Adversarial Robustness • Model Evals • Alignment Faking • Model Organisms • Mechanistic Interpretability • Steganography • Jailbreaking • Multi-Agent Systems.

Wanna talk? Book a 1-on-1 or do it the old fashioned way: samuelgerrit.nellessen{at}gmail.com