I’m Maaz Karim. I work on applied machine learning systems, especially the part where models have to survive contact with real constraints: messy data, limited compute, unclear product boundaries, and the need to be maintained by someone later.
The work I enjoy most sits between experimentation and delivery. That includes training and tuning models, building small data and inference pipelines, profiling bottlenecks, and turning one-off prototypes into systems that are simple enough to reason about.
Lately I have been especially interested in multimodal and video understanding, reinforcement learning for sequential decision problems, and the engineering work around making ML systems cheaper, faster, and easier to operate.