Selected space
ML Systems
Training infrastructure, kernels, and distributed systems for ML.
The visible part of machine learning is the model. The part that makes it work at all is the infrastructure around it — the kernels, the memory layouts, the distributed loops, the serving stacks. This space digs into those layers, because understanding them is what separates shipping prototypes from shipping systems.
What you'll build here
- a custom GPU kernel and a benchmark that explains why it is fast
- autograd from scratch inside a tiny tensor library
- a training harness for a small model spread across two machines
- an inference server with real latency and throughput budgets
Who it's for
Engineers who treat a model like a black box and are trying to look inside — more interested in the data plane than the prompt.
Tools & stack:PythonCUDAPyTorchTriton
Want to build in ML Systems?
Applications for the founding cohort are open.