Shreyas Kulkarni

• Implemented Graph, DP, and Time series algorithms for processing terabytes of data on GPUs. (CUDA, Triton, Python) • Developed asystembasedon RAG for internal documentation, leveraging LLAMA-2 and llamaIndex for advanced search capabilities. • Improved the throughput and latency of OSS LLMs for internal endpoints using continuous batching and paged attention (VLLM, Ray) • Researched on various decoding strategies to improve inference latency and throughput for LLMs while maintaining output quality. • Assisted graduate students in their HPC related projects and helped in maintenance of HPC cluster with more than 24000 CPU cores

• Led the architectural design and development of No-Code AutoML platform which serves and train’s 100s of models in production. • Reducedtheinfra costs by 22% by optimizing the inference graph of distributed document processing engine. (Ray, Kubernetes, GPT) • Improved the scalability of serving infrastructure and optimized models for latency and memory footprint. (Ray serve, Golang,gRPC) • Reducedtheinference latency of translation API by 37% using WebSockets andmodelleveloptimizations (Flask, Transformers,CUDA) • Built data pipelines to process terabytes of financial data for downstream ML tasks.(Ray, Spark) • Ledresearch initiatives in Text to SQL algorithms, exploring open-source datasets and SOTA models.(GPT,WikiSQL, Picard, RATSQL) • Developed a system that transforms hand-drawn form designs into fully functional backend applications. (Mask RCNN, YOLO, Django)

Wrote custom cuda kernels to reduce the latency ML algorithm for tone analysis and PI information detection in text. (NLP,CUDA) • Developed CUDA kernels to efficiently blur video streams in real-time, optimizing performance for live communication applications • Reduced the data retrieval time by 31% by optimizing the database using indexes and normalization techniques. (RDMBS, Postgres) • Experimented text generation using GPT to automatically draft initial patient communication messages based on visit notes. • Trained ML models for tasks like NER, sentiment analysis, text classification etc. (Python, NLP) • Developed scalable APIs for chat system, ensuring secure and accessible patient communication (Django, Mongo, AWS S3, ECR)