Pratik Jaiswal

Led development of an AI-driven Wealth Management Advisor collaborating with product managers, financial analysts, and data engineers to build personalized portfolio recommendations. Managed data collection, preprocessing, and storage on AWS S3, ensuring 98% data completeness. Developed reinforcement learning models using TensorFlow and OpenAI Gym to optimize portfolio rebalancing for maximal returns with high accuracy (F1-score 0.87). Applied feature engineering techniques including segmentation and risk factor encoding, improving model AUC by 12%. Tuned hyperparameters through Bayesian optimization and conducted extensive A/B testing leading to a 25% increase in adoption and client satisfaction. Containerized and deployed models on Kubernetes with RESTful API integration into digital banking platforms, ensuring compliance and real-time monitoring.

Implemented a real-time fraud detection system using Graph Neural Networks with cross-functional teams in data science, fraud risk, and compliance. Developed scalable data pipelines using Azure Event Hubs and Apache Spark on Azure Databricks, processing over 60 million daily transactions with 99.8% usable data. Built custom Graph Attention Network models utilizing PyTorch Geometric and NetworkX, deployed via Azure Machine Learning pipelines, resulting in a 38% detection accuracy improvement. Applied feature engineering including edge scoring and time-window aggregation, reducing feature space by 30% with PCA while preserving 92% information. Conducted model validation with stratified k-fold and A/B testing, boosting precision by 42% and reducing false positives by 22%. Managed deployment containerization on Azure Kubernetes Service and real-time monitoring with Prometheus and Grafana.

Collaborated with Compliance, Risk, and Engineering teams to define document verification requirements for KYC workflows, aligning with regulatory standards. Curated multilingual, low-quality datasets from PAN, Aadhaar, and Passport documents and built preprocessing pipelines with OpenCV and Tesseract for layout detection and text extraction. Developed CNN-RNN models in TensorFlow and Keras for document type classification, achieving a macro F1-score of 91%. Integrated models into Citi’s KYC system using Docker and REST APIs, creating a Streamlit UI for internal testing that reduced customer onboarding verification time by 70% across key regions.