Pratyush Chatterjee

Summary

Investigated cultural sensitivity in small-parameter large language models (LLMs), focusing on mitigating harmful cultural outputs. Fine-tuned LLMs, including Mistral-v0.2 (7B), Llama-2 (7B, 13B), Vicuna (13B), and Phi (4B), using advanced alignment techniques such as Direct Preference Optimization (DPO) and Offline Reward-based Preference Optimization (ORPO), achieving up to a 70% reduction in harmful outputs. Utilized Captum, an explainability library, to interpret model decisions and understand attention distributions influencing harmful responses.

Summary

Developed a multi-agent GenAl system to detect and support emotionally distressed employees through adaptive, empathetic conversations. Orchestrated autonomous agents using LangGraph and Agno, enabling context-aware, emotion-driven interactions powered by GPT-40-mini and RAG. Implemented anomaly detection with Isolation Forest & LOF to flag at-risk employees for proactive intervention. Deployed full-stack solution with real-time reporting, HR escalation, and scalable infra using GCP, GKE, Vercel, and GitHub Actions.

Summary

Architected and implemented an Al-driven exam question generator leveraging LangGraph, OpenAI's GPT-40, and open-source LLMs to dynamically construct CUET exam papers. Integrated vector embeddings for semantic indexing and retrieval of historical exam content, enabling context-aware information extraction. Designed a scalable multi-agent architecture with dedicated agents for topic distribution, context synthesis, and question generation, ensuring domain-specific precision. Engineered a resilient RAG-based workflow with layered exception handling and graceful degradation, incorporating fallback mechanisms for uninterrupted exam paper creation.

Summary

Developed a robust stock market analysis and prediction model using Python, leveraging historical data and technical indicators to achieve an accuracy score of 0.81. Used yfinance library to retrieve historical stock data. Created and Leveraged various technical indicators like - SMA, EWMA, MACD, Bollinger Bands, RSI, and Stochastic Oscillator to develop features for stock market analysis, enabling informed trading decisions.

Summary

Predicted flight delays with CatBoost, achieving an R2 score of 0.918 after extensive hyperparameter tuning. Minimized flight delays by 53.55% at Atlanta Airport using a OR-based optimization model, incorporating constraints like airport capacity and timing adjustments, using the optimization software Gurobi. Processed and cleaned 1.3M flight records, integrating weather data via the Meteostat API and handling missing values with MICE imputation.