Feb 2021 – Present · 5 yrs 1 mo

• I have experience working on various open-source projects such as Torchbench and fair idea about DeepSpeed. In these projects, I measured the performance and functional aspects of PyTorch-based models and actively contributed by identifying and raising open-source defects.
• I closely collaborated on various Media data loaders, including SSD, RESNET50, UNET2D, and 3D. I took the responsibility of validating media data loader layers such as flip, brightness, contrast, saturation, hue, and also verified certain hardware IPs.
• I have enabled the integration of various datasets, including COCO, IMAGENET, KINETICS 400, ACTIVITY-NET, CHARDES, among others, to ensure a diverse range of data sources for model training and evaluation.
• I have developed use cases for multiple AI features, enhancing customer usability, stack generality, and improving overall performance. I have also focused on optimizing host utilization for scale scenarios.
• Furthermore, I have conducted decode fidelity analysis of data loaders to assess the effectiveness of different media hardware IPs in achieving model accuracy. This analysis helps in determining the appropriate hardware IP for next-generation HPC products.
• In addition, I possess knowledge and experience with tools such as Nighties, Jenkins, Jira, Kubvirts, and containers etc
• Machine/Deep Learning Frameworks: Pytorch, TensorFlow
• Python Libraries : scikit-learn, keras, pandas, mumpy, openAI
• highly intreseted in : DATA SCIENTITES, DL, LLM, GEN AI , PROMPT ENGINERING

Jul 2019 – Jan 2021 · 1 yr 6 mos

• I have worked on developing a high-quality and performant AI inference engine called SpringHill. In this role, my responsibilities included:
• Enabling Meta (Facebook) AI Suites, including benchmarks, on the Intel AI Inference Accelerator (SpringHill).
• Working on a Recommendation System based model, measuring its performance, validating the layers, and working on fusion scenarios.
• Performing functional and bit-exactness testing on Meta's real RecSys workloads.
• Enabling randomization for models, subgraphs, layer validation, and fusion scenarios like (conv+relu).
• Ensuring functional, numeric, and performance coverage of deep learning operators, including extending coverage of deep learning operator and subgraph suites using an in-house developed randomizer.