Overview
As a Staff/Principal-level Machine Learning Engineer and technical lead, I design, ship, and operate production ML and GenAI systems end-to-end, turning ambiguous goals into measurable roadmaps, reliable architectures, and scalable execution. I hold a Bachelor's Degree in Computer Science from Columbia University in the City of New York. I am a U.S. citizen and do not require visa sponsorship.
I specialize in building high-impact, reusable foundations that make multiple teams faster: modern retrieval and ranking stacks, LLM/VLM-powered assistants, and distributed agentic systems with strong evaluation and safety guardrails. My work emphasizes practical optimization-cost-aware model routing, caching and batching, efficient serving (CPU/GPU), and compression techniques such as distillation and quantization to meet strict p95/p99 latency, reliability, and cost constraints in real production environments.
I am known for cross-functional leadership and technical decision-making that scales: aligning product, data, platform/SRE, privacy/security, and research stakeholders through crisp design docs, clear success metrics, and fast feedback loops. I value inclusive, high-trust teams and bring a calm, metrics-driven approach to building systems that are secure, observable, and maintainable over the long term.
My focus is generative and agentic AI: retrieval + reranking, RAG, tool-use orchestration, evaluation harnesses, and safety/guardrails paired with systems-level optimization (batching/caching, quantization/distillation, GPU efficiency, and cost-aware routing) to hit strict latency, reliability, and cost targets.
TECHNICAL SKILLS
- Software
- Programming: Python, Java, C, C++, Rust, Ruby, SQL, NoSQL, RESTful APIs, GraphQL, unit and integration testing
- Application: React.js, JavaScript, TypeScript, Swift, HTML, CSS, Django, Flask, Node.js, Express, Selenium
- Cloud: GCP, AWS, Azure, Digital Ocean, Netlify
- Tools: Git, VIM, VS Code, Splunk, Confluence, Jira, Bitbucket, GitHub Actions, Docker, Kubernetes, Linux, Shell Scripting
- Machine Learning
- Deep Learning: Agentic AI, PyTorch, TensorFlow, Keras, ML Recommender and Ranking Systems, Large Language Models (LLMs), Generative AI, RAG, LangChain, Multimodal Learning, Transformers, BERT, T5, Scikit-learn, NLP, NLTK, Knowledge Graphs
- Data: Pandas, Numpy, SciPy, PyTest, Spark, Hadoop, MapReduce, Tableau, Avro, Parquet, Data Parallelism, Model Parallelism, Hybrid Parallelism, Quantization
- MLOps: Ray, Slurm, Airflow, MLFlow, AutoML, Continuous ML, YARN, Kubeflow, Jenkins, Argo, CircleCI, GPU Scaling, vLLM, Distillation
- ML Systems: GRPO, SWiRL, DPO, PPO, Kafka, Zookeeper, ETCD, SHAP, LIME, NVIDIA NeMo, NVIDIA Inference Server, CUDA
Education
Columbia University
Bachelor's Degree, Computer Science (May 2020)
Coursework
- Artificial Intelligence with Python
- Natural Language Processing with Python
- Advanced Programming with C/C++
- Algorithmic Trading with Python (audit)
- Data Structures with Java
- Cloud Computing and Big Data in AWS, GCP, and Azure with Python, JavaScript, HTML/CSS
- Introduction to Cryptography
- Fundamentals of Computer Systems
- Linear Algebra
- Building a Technology Startup
- Computer Science Theory
Stanford University
Certificate, Machine Learning Specialization in Supervised, Unsupervised, and Advanced ML Algorithms (April 2023)
Coursework
- Supervised Learning: Regression and Classification
- Build machine learning models in Python using popular machine learning libraries NumPy & scikit-learn
- Build & train supervised machine learning models for prediction & binary classification tasks, including linear regression & logistic regression
- Unsupervised Learning: Clustering, Anomaly Detection, Recommender Systems, Deep Reinforcement Learning, Collaborative Filtering, Content-Based Deep Learning
- Use unsupervised learning techniques for unsupervised learning: including clustering and anomaly detection
- Build recommender systems with a collaborative filtering approach and a content-based deep learning method
- Build a deep reinforcement learning model
- Advanced Machine Learning Algorithms: Multi-Class Classification in Neural Networks with TensorFlow, Best Practices in Machine Learning Development, Random Forests, Boosted Trees, Regression Trees, XGBoost
- Build and train a neural network with TensorFlow to perform multi-class classification
- Apply best practices for machine learning development so that your models generalize to data and tasks in the real world
- Build and use decision trees and tree ensemble methods, including random forests and boosted trees
Experience
NVIDIA
Senior/Staff Machine Learning Engineer (March 2026 - Present) Boston, Massachusetts - Contract
-
Architected and shipped a Ray Jobs API distributed execution backend spanning three production codebases (NVIDIA NeMo-RL, NeMo-Skills, and the nvflow orchestration platform), enabling SFT and GRPO reinforcement-learning post-training to run on customer-managed Ray clusters for the first time (previously Slurm-only), unblocking a top-tier global investment bank's regulated, air-gapped on-prem deployment of NVIDIA's GenAI post-training stack
-
Proved the full SFT to GRPO RLVR pipeline end-to-end on multi-node [N]x H100/H200 GPUs - tensor-parallel policy, colocated vLLM generation, DTensor workers, and an LLM-as-judge reward loop - sustaining ~93 TFLOPS/GPU and landing trained Qwen3-class checkpoints with 100% unattended automation across a heterogeneous multi-cluster topology
-
Designed a heterogeneous multi-cluster Ray architecture (independent Python-runtime clusters) with automatic per-stage GPU/CPU cross-cluster routing, letting a single command fan an end-to-end RL workflow across clusters and cutting manual orchestration steps by 68%
-
Drove the open-source upstream contribution of the Ray execution backend into NVIDIA's NeMo-Skills project, clearing maintainer review with a guaranteed zero-behavior-change default path, adversarial multi-agent test coverage, and hundreds of regression tests proving byte-identical Slurm command emission - protecting 82+ existing production deployments from regression
-
Eliminated 100% of manual intervention from the RL run loop by engineering a shared-filesystem serve-reaping protocol and process-group GPU reclamation and text diffusion decoding, collapsing a multi-step babysat process into a single unattended run and recovering ~74 GB of leaked GPU memory per failed attempt that previously OOM'd downstream training
-
Led security hardening of air-gapped enterprise containers to pass the customer's vulnerability scan gate, driving fix-available High/Critical CVEs from [~437] to 0, filing [~92] upstream security bugs, and standing up a repeatable scan to remediate to promote pipeline producing signed, digest-pinned images compliant with the bank's air-gap and regulatory requirements
-
Validated cross-scheduler model-quality parity, reproducing evaluation benchmarks within +/- 2.5% pass@1 of the established baseline, and reduced judge-induced rollout failures by tuning rollout concurrency (24 to 4) to eliminate rate-limit storms
-
Authored cold-start self-serve documentation and troubleshooting runbooks enabling the customer to run the platform end-to-end without NVIDIA support, accelerating the experimental to production transition and defining the Slurm-compatibility regression bar plus automated NSPECT/Trivy release gating
Amazon
Tech Lead Machine Learning Engineer in Artificial General Intelligence Customization (AGI-C) Team (January 2025 - January 2026) Boston, Massachusetts - Full-time
-
Integrated GRPO, DPO, SWiRL, and RLVR into a unified RLHF stack to align LLM-generated inverse design code with verified metamaterial ground truth, enabling zero-shot generalization, physics-constrained reasoning, and unsupervised fine-tuning of instruction-following LLMs used by MIT researchers to accelerate synthesis validation for over 3 million material candidates
-
Engineered a batched, parallelized, and highly distributed reward evaluation pipeline across the 800TB MetaGen materials database, reducing runtime from 82 hours to 57 seconds for over 10 million completions, with >98% throughput efficiency using PyTorch, HuggingFace trl, and CUDA-aware sharded evaluation on multi-node clusters
-
Achieved >5000 times speedup in inference-time reward computation with >92.3% top-1 structural match accuracy using RLVR-based relative scoring, enabling scalable RLHF-style fine-tuning on commodity 16GB VRAM hardware via LoRA and 4-bit quantization, supporting models from 350M to 1.3B parameters
-
Developed a high-fidelity LLM evaluation framework for multi-turn scientific reasoning and code trace validation, analyzing over 2.1 billion tokens across 12 reasoning task types
- Utilized Hopfield episodic memory layers and text diffusion decoding to enforce logical and unit-consistent reasoning chains and Tree of Thought reasoning based prompt engineering techniques, boosting multi-hop pass rate by 94% and trace accuracy by 88%
-
Improved inverse design success rate from 46% to 97%, while reducing GPU-hour cost per training cycle by ~98%, demonstrating the real-world feasibility of modern RLHF and alignment techniques in high-throughput scientific GenAI pipelines
-
Enabled unsupervised fine-tuning workflows across domains, using self-consistency checks, rule-based constraints, and reward function introspection to curate alignment signals without human labeling—automating preference modeling for LLM self-alignment at scale
-
Deployed Self-Adapting Language Models to dynamically adjust reasoning behavior and output formatting based on prompt context, reducing domain-specific hallucinations by 87% and improving structured generation pass@1 by 83% in physics and materials applications