Performance Tuning Developer in New York, NY, United States

I was integral to the development of a model for fine-grained image recognition across several business verticals.

My contributions included writing a training system in Python that could effectively load images fast enough to maintain high GPU utilization during distributed model training runs while maintaining high developer flexibility, developing a novel way to aggregate contrastive loss pairs across GPUs that improved model accuracy, and catching several discrepancies between test systems and production which impacted the final accuracy of the model.

I contributed accuracy and speed improvements to a neural network used for duplicate image detection, including comparisons between a variety of different architectures and quantization into int8. I was also involved in the deployment of this model to production traffic, where I proactively caught a subtle issue that caused early load testing results to underestimate resource usage by order of magnitude and also managed to track down a subtle bug related to receiving different results for a fraction of a percent of production traffic which ended up being related to differences in CPU microarchitectures in our server cluster.