projects

A growing collection of my research and software contributions!

research
project thumbnail

Attention-Aware DPO for Reducing Hallucinations in Multi-Image QA

We introduce an attention-aware, multi-image augmented preference alignment method that improves accuracy by 8.5%, and further enhance inference-time alignment through adaptive attention scaling, yielding a 10% performance gain over the base model.

May 2025
DPO Multimodal VQA
In preparation
project thumbnail

Do retrieval heads speak the same language?

This project analyzes retrieval heads in multilingual LLMs using Needle-in-a-Haystack tasks across English, German, and Chinese. We find that strong retrieval heads are largely language-agnostic and critical for performance. Masking them leads to significant accuracy drops, offering insights for optimizing KV caching and multilingual model efficiency.

Mar 2025
NLP Interpretibility Causal Analysis
In preparation
project thumbnail

Towards synthetic data augmentation for Lecture Slide Understanding

SynSlideGen is a synthetic data generation pipeline that creates realistic, annotated lecture slides through an LLM-powered generation pipeline. Designed to support tasks like slide element detection and retrieval, it leverages structured text to generate diverse layouts and semantic content. SynSlideGen addresses the scarcity of annotated educational data through scalable automation.

Dec 2024
Document AI Computer Vision Synthetic Data
project thumbnail

ClinicalML - A medically interpretable ML pipeline for clinical outcome prediction using MIMIC-III discharge summaries

ClinicalML is an interpretable machine learning pipeline for predicting ICU patient outcomes—mortality and length of stay—using admission-time clinical notes. It extracts disease and drug entities, reduces dimensionality with BioClinicalBERT embeddings, and applies traditional ML models. ClinicalML performs comparably to BERT-based models while offering greater transparency and clinical trust.

Nov 2024
ClinicalML EHR NLP

software