Boston, MA
Principal Machine Learning Engineer · Cerby
Principal ML Engineer leading foundation model development at Cerby for identity and access management automation. My work spans building and fine-tuning large language models, reinforcement learning, and geometric deep learning. Currently exploring mechanistic interpretability, understanding how LLMs arrive at their decisions. Cross-domain experience across FinTech, Life Sciences, Shipping, and Utility sectors.
ICMLA 2024
Graph machine learning paper using Graph Neural Networks with attention-based residual connections to detect fraud in financial transactions. Outperforms state-of-the-art algorithms on benchmark datasets.
ICMLA 2024
Analyzes the impact of human-in-the-loop feedback on financial transactions and proposes a novel algorithm that propagates feedback signal further into the graph.
CIKM 2024 RAGEnt Workshop
Foundation model trained on Symbolic Regression to generate rule sets for financial fraud detection. Rule sets run 60x faster than other state-of-the-art methods.
AAMAS 2023 · with Prof. Karl Lieberherr, Ruiyang Xu
Step-MCTS: subnet structures within the complete network, each simulating a tree that provides lookahead for exploration. Optimizes training times of state-of-the-art AlphaZero MPV-MCTS.
Master's Thesis, Northeastern University · Advisor: Prof. Karl Lieberherr | Reader: Prof. Christopher Amato
Meta MPV-MCTS and Dual MCTS AlphaZero: outperform vanilla AlphaZero on combinatorial games using limited training budgets. Implemented efficient training strategies for asymmetric games.
arXiv · with Prof. Karl Lieberherr, Ruiyang Xu
Novel algorithm using a substructure within the neural network to achieve balanced accelerated learning. An intermediate policy is used to optimize the final policy that determines agent behavior.
arXiv · with Prof. Karl Lieberherr, Ruiyang Xu
Evaluates first-order logic using game-theoretic semantics by dynamically mapping to a two-player game (logic game).
Led the design of AI-driven workflow code generation using transformer-based language models and graph machine learning, modeling user interactions, DOM structure, and execution flows as graphs to generate, validate, and evolve CHAOS automations. Architected multimodal ML systems leveraging Scout and Replayer data (DOM, rrweb events, screenshots, videos, and failure traces) to power error classification, failure clustering, and drift detection in production workflows. Built and owned scalable end-to-end ML infrastructure on AWS, including data pipelines, SageMaker training pipelines, model deployment, and monitoring, enabling reliable iteration toward self-healing automations.
Led design and deployment of graph-based fraud detection systems using GNNs and custom graph algorithms, including JA-GNN which outperformed state-of-the-art methods on proprietary and public datasets, improving risk attribution by over 20% across multi-block transaction networks. Built distributed graph pipelines with PySpark, DGL, Neo4j, and Redis, including a custom in-memory graph library enabling real-time subgraph extraction, similarity scoring, and risk propagation.
Engineered scalable data pipelines for temporal and graph-based feature generation over billions of transactions, enabling efficient training and inference for large-scale fraud models. Delivered end-to-end ML solutions on Azure, containerized with Kubernetes (AKS), integrating graph outputs into downstream risk models and real-time scoring workflows. Collaborated with investigators and cross-functional teams to convert user stories into graph-native features and deployed models using open-source ML frameworks like PyTorch and DGL.
Built and optimized end-to-end ML pipelines to reduce liquid chromatograph equilibration time by 30%, employing advanced signal decomposition on time-series sensor data. Integrated AWS Redshift with PySpark-based ETL pipelines and deployed a Streamlit application on EC2, streamlining real-time data ingestion and SME collaboration for model lifecycle management.
Teaching Assistant for CS4100: Artificial Intelligence (Prof. Christopher Amato) and CS5100: Artificial Intelligence (Prof. Stacy Marcella).
Developed HPC-ready genetic optimization models for vessel routing, integrating climatic and onboard sensor data to achieve a 14% improvement in fuel efficiency. Implemented a Flask-based web app transitioning vessels from HSFO to VLSFO, containerized and deployed on Kubernetes for real-time, high-performance fuel management. Partnered with operations and technology teams to develop distributed, data-driven solutions using PySpark pipelines, optimizing maritime logistics and reducing operational costs.
Built and deployed predictive pricing models using Spark MLlib on high-volume utility data, enabling real-time rate adjustments for electric and gas clients. Collaborated with data engineers and domain experts to streamline data pipelines and integrate open-source ML frameworks for more efficient model experimentation.
Master of Science, Data Science (Thesis Track)
Thesis: Optimizations over Neural MCTS for Combinatorial Games (Advisor: Karl Lieberherr, Reader: Chris Amato). Coursework: Algorithms, Supervised/Unsupervised Machine Learning, Information Retrieval, Statistics.
GPA: 3.90
Cross-Registration from NEU
Coursework: Deep Reinforcement Learning
GPA: 4.00
Bachelor of Engineering, Electronics
Coursework: Object Oriented Programming, Data Structures, Digital Signal Processing
First Class with Distinction
Visiting Lecturer, University of Pune
Visiting Lecturer, University of Pune
Multidisciplinary Workshop on Stochastic Modelling and Monte Carlo Tree Search, TU Munich
Visiting Lecturer, University of Pune
Deep Learning, NLP
BioBERT-based models for extracting NERs from EHR data and a market-basket model for Relationship Extraction (RE) between extracted NERs to produce knowledge graphs.
Advisor: Prof. Karl Lieberherr
Implemented ideas from Discovering Reinforcement Learning Algorithms (Oh et al.) using the backward LSTM approach. Tested on Nim, Othello, Connect4.
Python, Dash, Heroku, Statsmodels, Plotly
Web application that lets the user upload a dataset and perform analysis or modeling using various machine learning algorithms.
Computer Vision, TensorFlow
Translates human faces across domains (expressions, color, gender, age) using the GAN configuration from StarGAN (Choi et al.).
Advisor: Prof. Karl Lieberherr
Multiple policy-value networks combined with the metacontroller from Deepmind's Agent 57 to achieve balanced learning in AlphaZero for combinatorial gaming problems.