GPU Systems & Inference Optimization | LLM Production Systems | Published ML Research
Production-grade benchmark comparing custom Triton attention kernels against FlashAttention-2 (CUDA) across both vLLM and SGLang inference engines under real concurrent load (1-32 concurrent users). Measures TTFT, token throughput, P50/P95/P99 latency, and GPU SM utilization. Demonstrates 12.3× throughput gains (573K → 6.03M tok/s) and 99.7% memory reduction vs vanilla attention on NVIDIA L4. Real hardware measurements on 4× A30 with reproducible benchmarks.
High-performance GPU kernels written from scratch in Triton and CUDA C++: (1) Flash Attention forward pass with online softmax recurrence, O(N·D) HBM vs O(N²), supports causal masking. (2) int8 GEMM + fused dequantization using DP4A, achieving 105–120% of fp16 cuBLAS throughput. (3) Fused RMSNorm+Linear (1.41–1.48× speedup). Benchmarked on NVIDIA A30 with 11/11 correctness tests passing. Beats torch SDPA at seqlen≥512.
Implemented lossless speculative decoding for SGLang on 4× NVIDIA A30 cluster. Custom draft/verify architecture with RadixAttention-safe provisional KV cache (insert/commit/evict operations). Accept-reject sampling verified mathematically lossless. Achieves 2.5–3.2× measured speedup on Llama-3-8B with TinyLlama-1.1B draft (K=4). Full unit test suite: 22/22 tests passing. Production-ready with real hardware measurements and reproducible benchmarks.
Two production-grade systems: (1) Custom CUDA C++ scaled-dot-product attention kernel with tiled QKᵀ, numerically stable softmax, and pybind11 PyTorch binding, 5.64x faster than PyTorch at N=32, correctness verified at max_diff < 1e-7. (2) GPT-2 ported to AWS Inferentia via neuronx-cc 3-step pipeline achieving 3.7x speedup (45ms to 12ms, 1,800 to 6,700 tokens/sec). Deployed as FastAPI REST endpoint on HuggingFace Spaces.
Supervised fine-tuning of Qwen2.5-7B using LoRA (r=8, alpha=16) on UltraFeedback, training only 0.5% of parameters (35M of 7B) with QLoRA 4-bit quantization and FP16 mixed precision. Achieved 17% training loss reduction (1.412 to 1.176) and 0.855 BERTScore in 30 minutes on a T4 GPU. Model merged and deployed to HuggingFace Hub.
Benchmarking framework comparing 4 attention implementations on NVIDIA L4 (seq len 1024, batch 32): FlashAttention-2 achieves 12.3x throughput (573K to 6.03M tok/s) and 99.7% memory reduction (12,582 MB to 38 MB) vs vanilla. Includes batch-size auto-tuner using binary search and ONNX/TensorRT export benchmarks. Key finding: algorithm-level optimization (FlashAttention) outperforms hardware-level optimization (TensorRT) by 6x for attention operations.
Multi-strategy reasoning system implementing 4 research papers: Chain-of-Thought with self-consistency voting (3 independent paths), Tree-of-Thoughts with beam search (width=3, depth=3), ReAct with real-time web search via Tavily API and ChromaDB vector memory, and Multi-Agent collaboration (Planner, Worker, Critic). LLM-based auto-classifier routes each query to the optimal strategy. Rate-limited (10/min, 100/day), with real-time streaming. Built with Groq LLM, deployed on HuggingFace Spaces.
Latest 15 repositories (auto-updated, 1hr cache):
Loading latest repos from GitHub...
AI Solutions Engineer Intern
Mar 2026 - May 2026 • New York, NY (Remote)
End-to-end LLM deployment across 6+ live hospital systems (Medstar, Mount Sinai, UFL). Implemented clinical LLM post-training pipelines (SFT → DPO) for medical question answering on real FHIR/EPIC data. Deployed and maintained inference endpoints on AWS with real-time latency monitoring. Configured domain-specific prompting for radiologist report parsing and automated EHR data extraction. HIPAA-compliant: built DPIAs, security assessments, and bidirectional EPIC integration documentation.
Machine Learning Engineer Intern
Dec 2025 - Feb 2026 • United States
Built computer vision pipelines for intra-oral image analysis across 6+ clinical categories (gingivitis staging, plaque detection, recession classification) on a 50K+ labeled dataset. Engineered automated retraining pipeline on AWS SageMaker incorporating dentist-corrected labels, improving model accuracy iteratively across production inference endpoints.
Graduate Research Assistant
Jun 2025 - May 2026 • Arlington, TX
Built TopGPT, a full-stack LLM application fine-tuned on 3+ textbooks with RAG over 1,000+ research paper chunks stored in Pinecone on AWS. Led CTMap, an LLM fine-tuning pipeline for mmWave 6G path planning, resulting in two accepted IEEE publications: IEEE ICC 2026 (conference) and IEEE OJ-COMS 2026 (journal). Built SFT pipelines encoding Sionna channel maps and OpenStreetMap graphs into transformer-readable formats.
Graduate Teaching Assistant
Aug 2024 - May 2025 • Arlington, TX
Supported graduate courses in Numerical Methods for 50+ students, assisting with algorithmic problem solving, optimization, and computational modeling. Concurrently developed CTMap, an LLM-enabled 6G path planning system accepted at IEEE ICC 2026 and extended to a journal paper at IEEE OJ-COMS 2026, fine-tuning LLMs on Dijkstra-generated coordinate paths applied to Sionna wireless network simulation outputs.
Senior Software Engineer → Software Engineer
Jun 2019 - May 2023 • 4 years • Chennai
Designed and owned Java-based distributed data processing services handling millions of records daily across production systems serving 10+ enterprise clients. Led system design and architecture reviews for data-intensive microservices. Built fault-tolerant backend pipelines with high availability and reduced processing latency by 40% through service refactoring.
Efficiency–Equity Trade-Off Characterization in Hierarchical Traffic Congestion Pricing
Structured efficiency–equity trade-off analysis within a hierarchical Stackelberg congestion pricing framework. Demonstrates that modest incorporation of equity considerations can meaningfully reduce travel cost disparities without significant congestion degradation, providing quantitative guidance for equity-aware congestion pricing design.
Digital Twin–Guided AI Path Planning for Connectivity-Aware Mobility
Comprehensive digital twin–guided AI framework for connectivity-aware mobility in 6G networks. Uses Sionna ray-tracing simulations to train AI agents that predict link quality along candidate routes and select paths maximizing sustained mmWave wireless connectivity.
CTMap: LLM-Enabled Connectivity-Aware Path Planning for mmWave 6G Networks
Fine-tuned LLMs on Dijkstra-generated coordinate paths from OpenStreetMap, applied to Sionna 6G wireless simulation outputs to produce real-time optimal paths for mmWave connectivity-aware routing.
Online · Powered by Claude
