Carnegie Mellon University
M.S. in Computational Data Science
Coursework includes Search Engines, Large Language Models Applications, Cloud Computing, and Machine Learning Systems.
CMU · Carnegie Mellon University · M.S. Computational Data Science
Sai Gopal Reddy Kovvuri
I work on machine learning systems for LLM inference, compiler-backed kernels, information retrieval, and distributed ML infrastructure at CMU.
About
Researcher and engineer at the intersection of ML and systems.
I am Sai Gopal Reddy Kovvuri, a Master's student in Computational Data Science at Carnegie Mellon University (CMU), and a Research Assistant with the CX Group and a Research Assistant with the Catalyst Group. My current work focuses on compiler and inference components across open-source ML systems, including Apache TVM, MLC-LLM, WebLLM, and FlashInfer-Bench.
Previously, I was a Product Engineer at Juspay, where I built production payment systems and RAG-based developer tooling. I completed my B.Tech in Computer Science at Shiv Nadar University with high distinction, and my research has appeared at BMVC and NCC.
Education
Academic Background
Shiv Nadar University
B.Tech in Computer Science
Graduated with CGPA 9.12/10, high distinction, and four Dean's List honors.
Experience
Research and Engineering
Feb 2026 - Present
Research Assistant, CX Group
- Engineering a distributed web crawler with fine-tuned LLM raters to acquire high-quality pretraining data at scale, with RL and reward-model filtering.
Feb 2026 - Present
Research Assistant, Catalyst Group
- Built compiler and inference components across Apache TVM, MLC-LLM, and WebLLM to improve deployment performance and device compatibility.
- Extended FlashInfer-Bench for newer model architectures and real SGLang inference workloads, improving benchmarking coverage.
- Built FP16 GEMM kernels for NVIDIA Blackwell GPUs using TVM/TIR and Python DSLs for kernel development.
Jun 2024 - Jul 2025
Product Engineer, Juspay Technologies
- Built CodeGen, an internal RAG-based developer tool that grounded LLMs in the codebase and automated 28% of payment gateway integration work.
- Integrated six payment gateways into the payment orchestrator, maintaining critical payment logic and encryption workflows.
- Enabled On-Us transaction processing for HSBC, reducing per-transaction network fees through direct in-network routing.
Dec 2023 - May 2024
Product Engineer Intern, Juspay Technologies
- Used Kibana and structured logging to analyze transaction logs and surface Redis cache performance issues per API flow.
- Contributed to microservices handling 175M+ daily transactions by shipping requirements, resolving production issues, and improving reliability.
Jul 2023 - Aug 2023
Data Science Intern, Code for GovTech 2023
- Built an on-demand data generation and fine-tuning pipeline for Hugging Face models from natural-language user prompts.
- Applied Stanford NLP's Demonstrate-Search-Predict framework to improve responses for government scheme queries.
- Built fuzzy-matching retrieval that improved rural village document matching accuracy.
Work
Projects
Cloud Computing, CMU
Twitter Recommendation Engine
Architected a three-tier Go microservice with gRPC serving 1,000+ RPS, backed by Spark/Scala ETL, MySQL, and AWS infrastructure managed with Terraform, Helm, and Kubernetes.
Machine Learning Systems, CMU
Distributed Training Systems
Implemented 2D parallel training from scratch using MPI collectives and Megatron-style tensor parallel communication, with ZeRO Stage-3 parameter sharding.
Search Engines, CMU
QryEval
Built an end-to-end neural search and RAG engine using BM25, dense retrieval, BERT reranking, SVMRank, and pseudo-relevance feedback.
Blockchain Application
WarrantyNFT
Blockchain-based platform for NFT warranty cards, enabling issuance, transfer control, and warranty management.
View ProjectNLP Toolkit
AutoTuneNLP
A toolkit for seamless data generation and fine-tuning of NLP models, packed into a single convenient workflow.
View ProjectPublications
Peer-Reviewed Research
National Conference on Communications, 2025
Revisiting Subject-Action Relevance for Egocentric Activity Recognition
Reddy, K.S.G., Prabhakar, M., and Mukherjee, S. Dual-stream CNN-LSTM approach for egocentric activity recognition.
IEEE XploreBritish Machine Vision Conference, 2024
UnSeGArmaNet: Unsupervised Image Segmentation using Graph Neural Networks with Convolutional ARMA Filters
Reddy, K.S.G., Bodduluri, S., Adityaja, A.M., Shigwan, S., Kumar, N., and Mukherjee, S. Graph-neural unsupervised image segmentation with ARMA filters.
BMVC ProceedingsSkills
Technical Toolkit
Programming
Python, Go, C++
ML and Data
PyTorch, scikit-learn, Hugging Face, NumPy, Pandas, FAISS
Systems and Cloud
AWS, Azure, GCP, Docker, Kubernetes, Terraform, Helm
Databases and Tools
MySQL, PostgreSQL, MongoDB, Redis, Git, Kibana
Contact