I build GPU-accelerated, distributed systems for scientific computing and AI inference: MPI solvers that scale across hundreds of ranks, CUDA and Triton kernels tuned against hardware limits, and LLM and model-serving stacks measured on latency, throughput, and cost.

Currently at QubitSolve, engineering a hybrid quantum-classical CFD platform on AWS ParallelCluster, Slurm, MPI, and NVIDIA GPUs, plus its inference layer (vLLM, TensorRT, Triton Inference Server). Before that, I built high-availability backend and event-streaming systems at HPE and Nokia.

• HPC scaling: cut a lattice Boltzmann CFD timestep from 184 s to 0.16 s (~1,150x), with a hierarchical MPI decomposition scaling 145.7x at 243 ranks
• GPU kernels: custom CUDA C and Triton kernels at ~480-505 MLUPS on an NVIDIA L40S (~12x CuPy, 36-38% of peak DRAM bandwidth), profiled with Nsight Compute and Nsight Systems
• LLM inference: benchmarked vLLM and SGLang on Qwen3-8B. FP8 KV cache doubled KV capacity for 1.59x throughput with zero preemptions at 128 concurrent requests, and prefix caching delivered 2.6x on shared prompts
• Model serving: deployed a PyTorch CFD surrogate on Triton Inference Server with FP16 TensorRT (batch-64 engine latency 7.90 → 2.87 ms); rejected INT8 after rollout-error validation
• Reliability at load: held 99.9% availability and sub-200 ms p95 through 10x rollout bursts on an HPE GreenLake control plane
• Research: co-author, "Toward Quantum Advantage in Computational Fluid Dynamics: A Variational Lattice Boltzmann Approach," accepted to AIAA SciTech 2027

Languages
Python · C · C++ · Go · Bash

HPC & Parallel Computing
MPI · Slurm · AWS ParallelCluster · NCCL · strong scaling · checkpoint/restart · Qiskit Aer · cuQuantum/cuStateVec

GPU & Kernels
CUDA C · Triton kernels · Nsight Compute · Nsight Systems · CuPy · PyTorch

AI Inference & Serving
vLLM · SGLang · TensorRT · Triton Inference Server · FP8 quantization · KV-cache management · prefix caching · continuous batching · tensor parallelism

Distributed Systems & Cloud
Kafka · FastAPI · PostgreSQL · Redis · AWS · Docker · Kubernetes · Terraform · GitHub Actions

Observability & Performance
Prometheus · Grafana · OpenTelemetry · Nsight profiling · benchmarking

• High-Performance Computing: MPI decomposition, scaling studies, Slurm workload modeling, and cluster automation on AWS
• GPU Performance: kernel fusion, occupancy and memory-bandwidth tuning, GPU-resident execution, multi-GPU communication
• AI Inference: LLM serving engines, quantization (FP8/FP16/INT8), KV-cache capacity, batching, and latency/throughput trade-offs
• Model Serving: TensorRT engine builds, Triton Inference Server, accuracy gates before deployment
• Distributed Systems: high-availability control planes, Kafka pipelines, SLOs, and incident response

• Distributed Systems
• Backend & Real-Time Systems
• AI & RAG Systems
• System Design
• Low-Level Design
• Algorithms & Data Structures

Explore my repositories or the pinned projects below.