I'm Sarvesh Sundaram

An emotionally aware movie recommendation agent that goes beyond traditional genre-based recommendations by reasoning over a user's preferences, constraints, and current mood. The system combines LLM reasoning with retrieval, structured movie data, and persistent emotional feedback to generate personalized recommendations that adapt to user context.

I developed the agentic workflow using LangGraph and Llama 3, integrating RAG and FAISS-based vector search with the TMDB API to retrieve relevant movies from natural-language requests. I also built an automated evaluation framework using DeepEval and IntellAgent to compare the agent against simpler LLM recommendation baselines.

https://github.com/SMZCartoons/Movie-Oracle-LLM-Agent

The system improved recommendation relevance by 30% and emotional awareness by 40% over single-pass LLM baselines, demonstrating how agentic workflows combining retrieval, reasoning, and user context can substantially improve recommendation quality. The project also established a reusable evaluation framework for measuring multi-step AI agents beyond traditional accuracy metrics.

A research project exploring parameter-efficient fine-tuning and quantization for vision-language models performing multimodal mathematical reasoning. The project investigated whether specialized QLoRA adapters could improve reasoning accuracy while reducing the computational cost associated with adapting large multimodal models.

I fine-tuned Qwen2-VL using QLoRA on the Math-PUMA dataset and evaluated its performance on MathVista. I developed four specialized adapters using PCA and K-Means clustering to group mathematical problems by type, allowing different adapters to specialize in distinct reasoning tasks. I also built an LLM-as-a-Judge evaluation pipeline to compare adapter and quantization strategies.

https://github.com/SMZCartoons/QLoRA-FineTuning-in-VLMs

The specialized adapter approach improved multimodal reasoning accuracy by 51% over the initial fine-tuned baseline while relying on parameter-efficient QLoRA rather than full-model training. Across 1,000 evaluation problems, the project demonstrated how targeted model specialization and quantization can improve VLM performance while reducing the computational requirements of model adaptation.

An end-to-end machine learning and data visualization system for analyzing historical crime patterns across New York City and predicting potential crime hotspots. The system combines large-scale crime records with temporal and geographic information to model how crime risk changes across locations and over time.

I worked on the machine learning and data-processing pipeline, using XGBoost to model crime patterns and engineering temporal and geographic features for prediction. We integrated hotspot predictions into an interactive dashboard backed by Google BigQuery, allowing users to explore crime trends across locations, time periods, boroughs, and offense types.

https://github.com/SMZCartoons/NYC-Crime-Analysis-and-Hotspot-Prediction

The project demonstrated that combining historical crime patterns with temporal and geographic context could produce localized hotspot predictions beyond simple historical-density analysis. The resulting system transformed large-scale NYC crime data into an interactive predictive application that could support more data-driven exploration of public-safety trends and resource allocation.

A high-performance computing project focused on GPU-accelerating helPME, an open-source C++ library for Particle Mesh Ewald calculations used in particle simulations. The project extended the library's existing MPI-based parallel architecture with CUDA execution to investigate GPU acceleration for distributed scientific workloads.

I ported computational routines from C++ to CUDA and integrated GPU execution with the library's existing MPI-based distributed architecture. I also developed correctness tests and performance benchmarking infrastructure to evaluate strong and weak scaling across different combinations of MPI processes, compute nodes, and GPUs.

https://github.com/philiplpaterson/helpme-cuda

The project extended a production-scale scientific computing library with GPU acceleration while preserving compatibility with its distributed MPI architecture. Benchmarking across 1–125 MPI processes and up to 4 nodes and GPUs provided insight into how CUDA acceleration scales within multi-node scientific workloads while establishing infrastructure for validating the correctness and performance of future GPU optimizations.