Mamadou Wane

I’m an engineer building Python and C++ systems for LLM inference, reinforcement-learning analysis, and Linux performance. Previously, I served as a combat engineer in the U.S. Marine Corps.

Right now I’m building Rollout Diff, an open-source package that reads the rollouts an RL trainer already writes and flags which behavioral features shifted during training, with the evidence behind each. It groups related shifts into ranked events and lets you inspect the rollouts behind them. It grew out of What the reward curve hides, a research note on four public reward-hacking runs. A second note, The cost of fewer false alerts, measures what calibrating its detector costs in sensitivity. A third, When 428 events become an investigation problem, audits its report on one of those runs.

I’m also building IPE, a compact Qwen3 inference engine checked logit by logit against Hugging Face Transformers. Its GPU correctness tests now pass on NVIDIA hardware, and benchmarking against vLLM has started.

Before that I built TVC, a C++20 controller and Python simulator for one-axis thrust-vector control, and measured its 500 Hz loop on Linux. Outside my own projects, I fixed a clang-tidy false positive in LLVM and a permission gap in claude-mem (both merged).

You can reach me by email or on LinkedIn.