Shreyansh Kumar

Developed an AI-Powered Expense Intelligence System offering personalized spend insights, fraud detection, and budget optimization tools for corporate clients. Collaborated with cross-functional teams and data engineers to preprocess transactional data stored on AWS S3, achieving a dataset completeness of 98%. Built reinforcement learning models in TensorFlow and OpenAI Gym to optimize real-time budget recommendations, with an F1-score of 0.87 in fraud detection. Engineered feature pipelines including user segmentation and temporal spend pattern analysis, improving fraud AUC by 12%. Used AWS SageMaker for scalable training and Bayesian optimization for fine-tuning, leading to a 25% increase in usage of smart spend insights. Validated models for regulatory compliance and deployed them using Docker, Kubernetes, and Flask APIs for integration into web and mobile platforms.

Designed an AI-powered, multi-agent tutoring platform that replaces static study material with adaptive explanations, quizzes, and revision workflows, delivering a 30% improvement in learner outcomes and 35% higher student engagement. Architected a scalable retrieval and reasoning pipeline with recursive content chunking, optimized embeddings, and metadata-aware vector search, boosting answer relevance and reducing topic-level confusion by 42%. Built a feedback-driven prompt tuning and evaluation loop to adapt to repeated learner queries, minimizing teacher intervention while aligning improvements with measurable academic outcomes. Developed an end-to-end intelligent document processing system combining CV enhancements, OCR, and hybrid rule-based + LLM extraction, achieving 100% processing success, 98% extraction accuracy gains, and a 90% reduction in API costs through intelligent method selection.

Led the development of a Python-based time series forecasting tool leveraging ARIMA, SARIMA, SARIMAX, VARMAX, and RNN models, reducing forecasting time by 70%. Automated model selection and hyperparameter tuning running 50+ iterations per dataset, accelerating delivery cycles from 3 weeks to 4 days for multiple projects. Improved model accuracy by 12% via iterative tuning and advanced feature engineering, supporting reliable predictions across industries. Enhanced client engagement by 15% through timely delivery of predictive insights to strengthen enterprise customer relationships.