Jun 2022 – May 2025 · 2 yrs 11 mos · Houston, Texas, United States · Hybrid

• Platform-Agnostic RAG System: Worked on a scalable RAG pipeline capable of retrieving upto 200M documents using Nomad/Kubernetes as an orchestration engine.
• Named Entity Recognition: Built high-performance pipelines in C++ for NER and PII detection, with a sub-millisecond inference latency on CPU.
• Checkpointing & Recovery in Distributed Training: A checkpointing routine for recovering failed ML training processes in a distributed cluster.
• Research Work:
• Automatic hyper-parameterization scheme for BOLT engine, achieving a 3-4x improvement in sparse inference accuracy.
• Designed a Random Unbiased Estimator (RUE) scheme for gradient compression to enable distributed training on low bandwidth.
• Created a BOLT-based Graph Neural Network beating SOTA benchmarks on datasets like Yelpchi.

May 2021 – Jul 2021 · 2 mos

• Identifying grammatical errors in text generated by Seq-2-Seq transformer models, using unsupervised and supervised training.
• Created heuristic functions and LM masking techniques for automatically generating synthetic data for training the model.

Jan 2021 – May 2021 · 4 mos

• Teaching assistant for the course MTL101: Linear Algebra and Differential Equations

Feb 2020 – May 2020 · 3 mos · Stockholm County, Sweden

• Designed and implemented a CNN for object detection and localization around self-driving cars, optimized for real-time inferencing with low latency.
• Integrated models like YOLO and RCNN into a ROS pipeline, handling data annotation and model training to enhance object detection capabilities in autonomous driving systems.