I build the infrastructure machine learning runs on — GPU kernels, compilers, and the harnesses that keep them honest — on top of seven years moving petabytes through distributed data systems.
Hyderabad, India. MS in Artificial Intelligence from Northeastern. Open to ML systems and distributed systems roles.
Hand-written CUDA and HVX kernels, each gated on a correctness check before its timing counts.
Inference serving and REST paths held under 100 ms for 10K concurrent users.
Streaming ingest at 70M records a day, under 2 s of lag.
2 TB of Spark batch inside a 6-hour SLA, and LLM fine-tuning ladders.
Built the backend and the multi-channel agent platform: REST on AWS EKS for 10K concurrent guests under 100 ms, with fail-closed LLM agents that cut in-stay response from 47 minutes to under one.
Owned streaming and batch: a Kafka and ClickHouse service ingesting 10K+ clickstream events per second, and a Spark system moving 2 TB inside a 6-hour SLA.
Medical equipment tracking on Kafka, Spring Boot, and Elasticsearch at 70M+ records a day under 2 s lag, plus a petabyte-scale migration to AWS at 80% lower cost.
Compilers, GPU kernels, and the harnesses that keep them honest. Everything below is public and runnable.
Correctness and timing harness for Triton kernels
A fast kernel is worthless if it is quietly wrong. Crucible verifies a kernel against an fp64 CPU reference at 95% confidence and ships as a zero-config pytest plugin, so the check runs wherever your tests already run. It has caught real defects: tail-dropping and contiguity assumptions failed 5 of 16 and 15 of 16 generated shapes, and every failure minimises to a reproducible seed you can paste into a bug report.
Compiler and kernel library for Qualcomm Hexagon NPUs
Compiles a vision transformer down to an explicit VTCM and DMA plan, then runs it on hand-written HVX and HMX kernels. Each kernel is gated on three things before it counts: correctness, cycle count, and vector use proven by disassembling the ELF rather than trusting the source. Targets Hexagon v75, the NPU in the Snapdragon 8 Gen 3.
GPU-accelerated pattern search library
Scans 10 GB+ files by chunking them through GPU memory automatically, behind a single API call that serves both in-memory buffers and files on disk. Supports regex wildcards, quantifiers, and character classes, with kernels tuned for Ampere.
Also on GitHub: fuseflow, an MLIR pass that fuses neural network operators; reduction kernels benchmarked against Triton, PyTorch, and CUB; mossformer2-denoise, speech denoising shipped in both PyTorch and Rust; openMed, medical imaging with GradCAM explanations; and Triton with TensorRT-LLM on Kubernetes.
Four wins, judged against open fields. Prize money and field size, as awarded.
Best in Enterprise Computer Vision
Winner
2nd Runner-Up
AI traffic signal control. A YOLO model reads vehicle density straight off live camera feeds, and a Deep Q-Learning agent retimes the signals from the queue lengths and speeds it sees — cutting average waiting time 40%, with a pre-emption path for ambulances and other priority vehicles.
Best mentor
Fine-tuned llava-1.5-7b, a 7-billion-parameter vision-language model, into a chat instruction model over a dataset of paired images and text. Trained with LoRA on an A100, with the base model quantised to 4-bit so the run fit on a single GPU.
Fine-tuned a transformer encoder with an MLP decoder into a semantic segmentation model over 18 labels and 17K samples. 3.7M parameters, mIoU 0.514. Weights are published on Hugging Face.
Also: pre-training versus fine-tuning, a controlled comparison of ResNet-18, VGG-19, and Inception V3 trained from scratch against the same architectures fine-tuned from ImageNet weights.