# Smit kumbhani > Machine Learning Engineer @Amazon | Ex-Research Intern @Zeblok (Post-Training) | RA @SBU (2D/3D Diffusion Models) | MS CS @SBU Location: New York, New York, United States Profile: https://flows.cv/smitkumbhani As a passionate AI/ML Engineer, I specialize in building scalable, intelligent systems at the intersection of foundational research and applied innovation. Currently driving improvements in Prime Video Search at Amazon, I bring deep experience in LLMs Post-Traininig, Agentic Systems, and Distributed AI traininng(e.g NCCL, FSDP/DDP) At Zeblok, I led efforts in RLHF, DPO, and Fully-Sharded Data Parallelism (FSDP) to scale SFT and RLSF on 70B models. I built a Chain-of-Thought (CoT)-driven synthetic data pipeline that cut generation costs by 12%, and designed a robust RAFT process for enhanced domain adaptation. Previously at Stony Brook University, I explored Diffusion Models for 2D/3D medical imaging, proposing a novel multi-conditioned latent refinement DDPM and advancing training-free style injection techniques for image synthesis. My work at SigmaRed focused on ML Monitoring, deploying real-time anomaly/drift detection for CV/NLP pipelines using MongoDB, FastAPI, and MLOps best practices. I also collaborated on XAI systems and engineered robust RESTful APIs for production ML. Tech Stack: vLLM, AKS, LoRA, FSDP, DDPM, CT-guided refinement, CoT prompting, FastAPI, Express.js, MongoDB, Kubernetes, MLFlow, Explainable AI, RAG, RLHF, DPO, SFT, RLSF. Driven by curiosity, I thrive in fast-paced, research-to-deployment environments building frontier AI solutions that scale. ## Work Experience ### Software Engineer, Machine Leaning @ Amazon Jan 2025 – Present | New York, New York, United States - Building Evals for search & browse Agents. - Working on Video Search and Ranking/relevance at Prime Video ### Research Engineer Intern @ Zeblok Computational Inc. Jan 2024 – Jan 2025 | New York, United States • Designed contrastive reward modeling and PPO-based alignment pipeline (RLSF) to ground responses of reasoning model in corpus human preference data, adapting RLHF for synthetic feedback alignment. • Integrated Fully-Sharded Data Parallelism(FSDP) for Supervised Fine-Tuning (SFT) and Reinforcement Learning with Synthetic Feedback (RLSF) to scale Post-training of 70B Model over distributed AI infra. • Orchestrated the scalable model serving of 405B LLM over multi-node Kubernetes clusters using Ray and vLLM, optimizing inference latency to handle high-volume of requests in production. • Enhanced domain-specific answering capabilities with reasoning to improve the knowledge depth of small language models for context-aware question-answering. • Engineered multi-phase CoT-based data generation pipeline for domain-specific Retrieval Augmented Fine-Tuning (RAFT), resulting in a 12% cost reduction in the synthetic data generation pipeline. ### Research Assistant (2D/3D Diffusion Model) @ Stony Brook University Jan 2023 – Jan 2024 | Stony Brook, New York, United States • Proposed multi-conditioned iterative latent variable refinement-based DDPM for image-to-image for CT Image generation. • Engineered novel training-free style injection method to significantly improve image-guided latent refinement in mammogram patch synthesis. • Dedicated to refining the noise scheduler for sampling procedures and mitigating signal-to-noise ratio to (0.12%) in the forward diffusion process. ### Machine Learning Engineer @ SigmaRed Technologies. #BeatAIBias Jan 2021 – Jan 2022 | Remote • Collaborated with the team of 4 mlops engineers and delivered end-to-end ML monitoring platform, including building feature engineering pipelines and performing inference of adversarial detection. • Built touchless ML model productionalization pipeline for real-time ML inference services, enhancing streamline model lifecycle management, versiong and experiment tracking, which improved deployment speed by 15% and optimised ML workflows in production environment. • Developed 5% enhanced LoRA-based contextual embedding model to improve drift detection algorithm to enable precise detection of data-distribution shifts in data embeddings. • Rolled updated version of feature store database and scalable data ingestion infrastructure using Apache-Kafka and pySpark, resulting in a 8% boost in high-performance data processing. ### AI/ML Intern @ SigmaRed Technologies. #BeatAIBias Jan 2021 – Jan 2021 | Remote ◦ Researched various Drifts and their detection algorithms. ◦ Developed end to end Adversarial and Drift detection mechanism and deployed it into the production pipeline using MLOps principles ### AI/ML Intern @ DST-SERB funded project SMART FOUNDDRY-2020 Jan 2020 – Jan 2020 | Marwadi University, Rajkot - Developed an Investment Casting Defect Analysis System using Machine learning and different evolution algorithms. ### Applied AI BootCamp @ Applied AI Course Jan 2019 – Jan 2020 | Bengaluru, Karnataka, India • Successfully completed an intensive 150+ hour training program in data science and machine learning, focusing on practical applications and hands-on learning. • Built proficiency in advanced computer vision techniques, machine learning algorithms, and deep learning methodologies through comprehensive projects and exercises. • Developed strong technical expertise in frameworks and libraries such as PyTorch, TensorFlow, Keras, NumPy, and NLTK, utilizing them to implement and optimize machine learning and natural language processing solutions. • Acquired the skills to leverage advanced computer vision, machine learning, large language models (LLMs), and deep learning to solve complex, real-world problems in a business context. ## Education ### Master's degree in Computer Science Stony Brook University ### Bachelor of Engineering in Computer Engineering Marwadi University ## Contact & Social - LinkedIn: https://linkedin.com/in/smit-kumbhani-44b07615a --- Source: https://flows.cv/smitkumbhani JSON Resume: https://flows.cv/smitkumbhani/resume.json Last updated: 2026-04-05