Varun Pradhan

• Developed a multi-agent communication framework for coordination between LLM-powered agents, simplifying the integration of new agents, and enabling direct context exchange, reducing redundant queries by ∼40%. • The framework serves as a knowledge base and starting point for a future AI agent marketplace. • Provided architectural insights for a customer-facing agentic app integrating agents from iFood and Despegar. • Conducted an exploratory evaluation comparing OpenAI Codex and Cursor, evaluating accuracy, latency, and developer workflow impact, informing selection for future use by the rest of the team

• Researched the feasibility of forecasting the head trajectories of the viewer in VR using user predictability metrics. • Discovered correlations between the entropy of user trajectories in VR and prediction errors through extensive data analysis. • Designed and evaluated models incorporating predictability metrics, demonstrating up to a 34% relative reduction in prediction error variability for videos with highly unpredictable user movement, indicating higher predictive stability. • This work resulted in the following publication: Pradhan, V., Rossi, S. & Cesar, P. Exploring Entropy-Based Solutions for Trajectory Prediction in Virtual Reality (MMVE 2025). • Participated in academic seminars and supported the VQEG Standardization Meeting on Immersive Communication Systems

• Explored sentiment analysis and sentence similarity techniques to assess their applicability in AI-driven conversational systems. • Implemented a weighted n-gram–based sentence similarity approach from prior research for internal benchmarking. • Evaluated pre-trained sentiment analysis models on the Stanford Sentiment Treebank (SST-5), identifying a transformer-based model achieving 49.27% accuracy as the most effective option relative to resources spent on fine-tuning.

Thesis: Pro-Nuance: English Pronunciation Trainer, DJ. Sanghvi College of Engineering and IIT-Bombay • Designed a Computer Assisted Pronunciation Training System that provides immediate, personalized feedback as opposed to the generalized learning approach of existing systems. • Implemented a probabilistic user similarity based algorithm that prioritizing words with a high likelihood of mispronunciation, based on shared user-error patterns, leading to more targeted and effective feedback in pilot evaluations.