shopify

applied machine learning engineering co-op

2026 · toronto, on · current

gpu training and serving work on shopify's recommendation models.

— made hstu training 24% faster by fusing layernorm and activation into triton kernels, at fixed batch and sequence lengths.
— cut peak gpu memory 18% with gradient checkpointing and tensorrt-llm inflight batching, at the same retrieval recall@10.
— scored 45 agent-eval tasks for quality and cuda latency, and isolated oom spikes during continuous batching.
— caught 2 regressions with reproducible gpu tests in metaflow-run ci/cd.