Software Engineer focused on high-performance systems, GPU computing, and local LLMs. I work at the intersection of low-level CUDA programming and modern ML — from kernel-level optimizations to memory-efficient LLM training and inference.
- Scalable backends — Spring Boot microservices on Kubernetes, Apache Spark, AWS EKS
- GPU-accelerated pipelines — CUDA kernels, cuPy/cuDF, Nsight/nvprof profiling, parallel algorithm design
- Local LLMs — LoRA/QLoRA fine-tuning on NVIDIA GPUs, exploring Apple MLX for on-device inference
- Distributed ML systems — multi-GPU training with NCCL, torch.distributed, mixed precision, gradient checkpointing
Languages C/C++ · CUDA · Python · Java · Bash
ML / DL PyTorch · TensorFlow · LoRA · HuggingFace · Apple MLX
GPU CUDA · cuPy · cuDF · Nsight · nvprof
Infra AWS EKS · Docker · Kubernetes · Terraform · Jenkins
Algorithms Numerical Optimization · Spatial DS · Graph Algorithms
Side-by-side benchmark of LLaMA training throughput and memory efficiency on Apple Silicon (MLX) vs NVIDIA CUDA GPUs. Explores unified memory advantages, framework differences, and practical trade-offs for local LLM development.
🏆 OpenAI Parameter Golf (contributor)
Competition to train the smallest LM that fits in 16MB. Explored aggressive quantization and model compression techniques on 8×H100 hardware.
Data harmonization pipeline unifying fragmented labor market datasets — normalization, schema standardization, and semantic enrichment using Python (Pandas, NumPy).
- M.S. Artificial Intelligence — San Jose State University Coursework: CUDA Architecture, GPU Programming, Deep Learning, Computer Vision
- B.Tech Computer Engineering — University of Mumbai
| Project | Result |
|---|---|
| CUDA preprocessing pipeline (160K+ records) | 45 hrs → 3 hrs (15× speedup) |
| LoRA fine-tuning on LLaMA (A100) | ~90% memory reduction, no quality loss |
| Locally hosted finetuned Llama 3 model on Apple Silicon | Hosted a server to process requests locally |
Currently exploring MLX for on-device inference