Building Open Posttraining System — reliable, distributed RL infrastructure for training swarms of agents.
I am building an open post-training framework dedicated to advancing open-weight models using reinforcement learning on complex coding and agentic tasks.
My work spans the full RL post-training stack—from distributed orchestration and sandboxed runtime environments to verifiers, inference, and model evaluation. Beyond infrastructure, I conduct controlled experiments to study training dynamics, inspect edge-case failures, and scale credit assignment over long execution trajectories.
On the systems side, I focus on end-to-end efficiency: writing custom kernels, optimizing multi-accelerator memory and communication overhead, and eliminating execution bottlenecks to maximize hardware utilization.
- Linkedin: @shaheennabi