Building scalable machine learning systems that are fast, reliable, and useful.
I'm Kushal, a Computer Science graduate student at the University of Rochester focused on applied AI systems, GPU-accelerated deep learning, and production-grade ML infrastructure.
I build end-to-end machine learning systems — from data pipelines and model training to scalable deployment.
My work focuses on building practical AI systems that are reliable, fast, and usable in real environments. I’ve worked across medical imaging, retrieval systems, and large-scale model training, combining deep learning with cloud infrastructure and GPU-accelerated workflows.
Recently, I’ve been building systems involving diffusion models, transformer architectures, and retrieval pipelines while integrating tools like PyTorch, FAISS, Spark, and AWS to support scalable ML experimentation and deployment.
What I'm currently exploring: combining large language models, retrieval systems, and generative models to create faster, more useful AI applications.
Fine-tuned Qwen2.5-14B-Instruct with LoRA (PEFT) for multi-class discourse classification on imbalanced datasets, raising accuracy from 76% to 89% and macro F1 from 0.63 to 0.81. Engineered rationale supervision and threshold tuning, boosting minority-class F1 by 18% and recall by 14% on noisy data.
Deployed a cloud-native ML pipeline on AWS (Lambda, Batch, EC2) with Docker, processing 2,000+ waveform samples. Engineered an S3 + DocumentDB layer indexing 10,000+ metadata entries for production-grade MLOps infrastructure.
Shipped an end-to-end video segmentation pipeline (MedSAM2 + CycleGAN) with a Gradio inference UI, automating annotation via MATLAB ROI labeling and eliminating ~70% of manual labeling. Built a Dice/IoU evaluation harness achieving a 0.85 Dice score.
Developed real-time weather forecasting models using SARIMAX, Random Forest, and Gradient Boosted Trees ensembles. Improved forecast accuracy by 21% and reduced launch delays by ~15% through systematic feature engineering.
If you're building products involving applied AI, generative models, or scalable ML systems, I’d be happy to connect.
Send a short note — I’ll get back quickly.