Projects

Selected Projects

Multimodal Retrieval-Augmented Generation for Visual Question Answering

Course project · ECSE-555 · 2025

A multimodal RAG system that combines SigLIP and CLIP-based encoders to retrieve both image and text documents for visual question answering.

• Benchmarked on TextVQA, VQAv2-Small, and a custom engineering QA dataset
• Evaluated with soft VQA accuracy, top-3 accuracy, and BLEU
• Studied retrieval-size ablations and the effect of captioning on retrieval quality
• Integrated Jensen–Shannon divergence for context filtering

Deep Learning-Based Classification of Dopamine Fluorescence

Course project · ECSE-552 · 2025 Developed deep learning models to automate analysis of dopamine fluorescence in brain slices from Parkinson’s disease mouse models treated with interferon gamma (IFNγ).

• Collaborated with neuroscience researchers to analyze fluorescence imaging data
• Implemented three models: feature-based FFNN, 1D CNN on frame-level signals, and 3D CNN on full video sequences
• Demonstrated improved performance with higher dimensional inputs using 3D CNN architectures
• Applied interpretability techniques to analyze the influence of brain slice identity and stimulation events on model predictions

Out-of-Distribution Image Detection and Segmentation

Research project · 2023–2024

Developed a zero-shot multimodal method for detecting out-of-distribution (OOD) objects in semantic segmentation tasks.

• Conducted a comparative analysis of modern OOD detection techniques for segmentation models
• Proposed a method that identifies anomalous objects using semantic similarity between predicted labels
• Designed the approach to generalize across domains without task-specific retraining
• Evaluated on the SMIYC and Road Anomaly benchmarks using pixel-level metrics (AuPRC, FPR95) and component-level metrics (Mean F1, PPV, sIoU)

Significant Wave Height Prediction with Machine Learning

Thesis project · 2022–2023

Developed machine learning models to predict ocean significant wave height from atmospheric and wind parameters.

• Curated and processed large-scale oceanographic datasets from 47 buoys across North America
• Trained XGBoost, LightGBM, ANN, and SNN models for regression
• Performed feature engineering and hyperparameter optimization
• Evaluated models using MSE, MAE, and R², achieving test R² > 0.90 on over 2M samples

Surface Defect Detection Using Representation Learning

Research mentorship · 2024–2025

Co-supervised a research project on automated fabric surface defect detection using representation learning techniques.

• Guided the development of supervised models using CNN and transformer-based architectures
• Evaluated unsupervised anomaly detection using EfficientAD
• Conducted experiments on the ZJU-Leper dataset for defect classification and segmentation
• Analyzed performance at both pixel and region levels, including qualitative and efficiency analysis