Examination of whether LLMs can maintain consistency over extended multiple text generation for 10 medical personas. 5 novel plausibility metrics proposed, and an ontology of common LLM errors.

Questo progetto di tesi affronta la possibilità di usare i Large Language Models (LLM), in particolare FlanT5-Large, ed il Reinforcement Learning (RL) per creare un NPC in grado di assistere un giocatore, suggerendo la prossima azione da intraprendere in un gioco a turni.

An open-source conversational AI assistant for rural & semi-urban India. Supports voice-based queries in Hindi, Punjabi, Tamil, Bengali and more. Converts speech to text (ASR), retrieves answers via RAG, translates between Indic ↔ English, and responds back with Indic TTS. Built with AI4Bharat models, Whisper, Flan-T5, IndicTrans2, and IndicLID.