Nitish Kumar Ranjan

YuVoice (Conversational AI BOT)

• Led the end-to-end development and deployment of a production-grade conversational AI Voicebot, managing a scalable GenAI platform that automates customer interactions and supports real-time decision workflows.
• Architected a stateful, multi-agent AI system using LangGraph and Model Context Protocol (MCP), effectively handling complex, multi-turn conversational workflows beyond rigid decision trees.
• Engineered an enterprise RAG pipeline, integrating vector search and proprietary knowledge bases to deliver highly grounded, context-aware responses.
• Built advanced orchestration and routing logic across multiple models (GPT via PTUs and Gemini), enabling dynamic, task-specific model selection to optimize for maximum accuracy, low latency, and cost-efficiency.
• Developed rigorous evaluation and monitoring frameworks encompassing prompt optimization, strict guardrails, and response quality metrics to ensure absolute reliability in a live production environment.

Next Best Action Model (Recommendation Model)

• Engineered a 'Next Best Action' recommendation model leveraging deep learning (LSTM and Functional APIs) to predict the optimal medium, time, and language for borrower outreach, significantly boosting repayment propensity.
• Utilized comprehensive EDA and feature engineering on historical campaign and Customer Input Parameter (CIP) data to generate sophisticated input sequences for targeted customer engagement.

Loan Propensity Prediction

• Designed and deployed predictive tree-based models (XGBoost/GBM) to forecast borrower repayment propensity across various delinquency stages (pre-due and post-due).
• Spearheaded extensive EDA and feature engineering pipelines to extract actionable signals from raw credit card datasets, enhancing baseline model performance.
• Implemented Decile Analysis to rigorously evaluate model precision and ensure accurate risk segmentation of the customer base.

Sentiment Analysis and NLP Model Development

• Designed a multi-class Sentiment Analysis model leveraging XGBoost and TF-IDF to accurately determine text polarity across diverse data segments.
• Utilized NLP techniques, including tokenization, lemmatization, and vectorization, to preprocess data for machine learning models.
• Delivered high-accuracy sentiment classification (positive, negative, neutral, mixed), enabling data-driven insights into customer behavior and intent.

Transaction Scoring Model

• Developed a monthly transaction scoring model for CredAvenue, incorporating parameters such as Credit Enhancement, Collection Efficiency, Current Rating, and Rating Changes.
• Implemented regression and weighted sum techniques to accurately score transactions based on key financial metrics.

Investor Preference Recommendation Algorithm

• Developed a recommendation algorithm for investors based on preferences such as revenue, ratings, EBITDA, PAT, financial ratios, sector interests, company vintage, and profitability.
• Integrated multiple data sources, including client financial data, Charges Data, and customer data from Redshift to generate a recommendation score tailored to investor profiles.

Term Loan Pricing Model

• Created a Term Loan Pricing Model employing advanced machine learning (Random Forest, XGBoost) to generate accurate pricing recommendations for institutional investors.
• Executed end-to-end data pipelines, including rigorous data cleaning, imputation, and multiple sampling techniques (Normal, Stratified, Cross-Validation) to guarantee robust and unbiased predictive modeling.