San Francisco Bay Area
As an Applied Researcher at eBay with over four years of experience, I specialize in machine learning and deep learning solutions to address fraud detection challenges. My work includes designing custom transformer architectures and multimodal models to enhance user behavior analysis and listing authenticity, contributing measurable improvements to fraud detection systems. Collaborating with cross-functional teams, I spearheaded the deployment of a billion-parameter model optimized for latency and scalability, enabling robust performance at scale. Prior roles, including as a Senior Data Scientist at PayPal, strengthened my expertise in language modeling, ensemble methods, and active learning strategies. These efforts resulted in enhanced fraud detection precision and substantial cost savings. With a mission to innovate through data-driven insights, I focus on leveraging distributed training, multimodal learning, and cutting-edge AI techniques to create impactful solutions that empower organizations to combat fraud effectively. Specialties: • Natural Language Processing , Multi-modal learning • Generative AI / Large Language Models : Transformers, BERT, GPT, LLMs, Distributed Training (DDP,FSDP), Ray • Cloud Computing - Microsoft Azure Batch Processing, AWS • Deep Learning - Neural Networks,Convolutional Neural Networks(CNN),Recurrent Neural Networks(RNN), Transformers • Word Embeddings - Transformer Encoders/Decoders, Word2Vec, Glove word Vectors • Machine Learning algorithms - Linear/Logistic Regression, Decision Trees(Random Forest, Bagging, Gradient Boosting - XGBoost/CatBoost) • Programming Skills - Python (numpy, pandas, scikit-learn), Pytorch, TensorFlow, SQL, Spark, HiveQL • Big Data Processing tools - Hadoop, Hive • Statistical Analysis • Predictive Modeling To further connect, i can be reached either through this profile or by email at [email protected]
Developed unsupervised representations of user page view sequences using custom transformer encoder and decoder architecture to represent behavior patterns between good and fraudulent users, thereby providing additional 5% lift to downstream models. • Architected a hybrid, two-stage scam detection pipeline combining a high-throughput SimHash/LSH engine with a semantic re-ranker, boosting scam detection recall from 0.65 to over 0.80 and protecting over $900M in targeted GMV. • Spearheaded the deployment of a 1B parameter model using DDP across 45 GPUs, optimizing for improved latency and scalability • Innovated a multimodal fused image and listing representation model using unsupervised contrastive anomaly loss, significantly improving the detection of fictitious listings and surpassing existing benchmarks • Enhanced topic intent model accuracy while reducing labeling costs by $1M through active learning and label augmentation techniques; utilized LoRA and Reinforcement learning for LLM fine-tuning • Developed a high-precision seller registration model, achieving a 92% precision rate and saving over $3M by mitigating fraudulent activities
• Engineered features for detecting gibberish and risky email patterns using language modeling and logistic regression, boosting fraud detection precision to 60% • Advanced PayPal’s fraud detection capabilities with ensemble models, saving over $1M through improved transaction scrutiny • Developed data exploration module for Tigress, in-house ML library, that performs necessary aggregations and visualizations to better understand fraud trends in the dataset, and also generates power point presentations
• Architected and implemented store location tagging process in a distributed processing environment using Microsoft Azure Batch service, leading to sharp reduction in processing time from 8 days to 3 hours, thus eliminating system maintenance as well as streamlining rapid deployment of data assets across the company • Developed and enhanced de-duping of store names and locations data sourced from multiple 3rd party merchant aggregators, thus improving the quality of match by 30% and ultimately generating more revenue for the company • Built a deep learning model using Word Embeddings and Neural Networks (LSTM) to predict clean merchant identifiers from merchant descriptions, thus achieving match accuracy of 96%; used Keras and Python for modeling • Built a Named Entity Recognition model using Spacy to extract clean merchant identifiers from unstructured data, thus automatically tagging merchants at scale and increasing revenue by 30%. • Liaised with senior leadership and sought feedback and input while solving complex business problems for clients