Alexander is a Columbia CS graduate student focused on machine learning systems, agentic AI infrastructure, and evaluation workflows. Among other topics, his writing explores model behavior, AI reliability, and interpretability. Before returning to CS, Alexander spent over a decade in fintech, analytics, product operations, and consulting, and he saw that the most common failure mode was poor cross-functional collaboration. This led him to develop his own local, scalable AI orchestration systems where agents can coordinate, delegate, execute, and review; suddenly workflows actually work.

Recent projects he led at Columbia include SmartGridBench, an open-source contribution to IBM’s AssetOpsBench for smart-grid transformer maintenance using multi-domain, tool-using industrial agents; and The Slippery Slope, an AI-safety study exploring reward-gaming in Gemma-2-2B through dose-controlled LoRA fine-tuning and leakage-aware activation probing. Across both projects, he used file-backed multi-agent coordination patterns that he engineered from scratch, with agents splitting implementation, review, evaluation, documentation, and final-deliverable auditing.

Outside the machine room, Alexander is a serial learner and adventurer: traveling, rock climbing, wandering cities by public transit, and disappearing into whatever subject has recently hijacked his curiosity. Other interests include music, theater, nature, and puzzles — the most interesting being the human condition. Definitely helps explain his curiosity for dramatic AI meltdowns, as a mirror of our own psychology.

GitHub: eggrollofchaos

LinkedIn: waximus