Projects

This paper describes a system which focuses on detecting hallucination spans in multilingual LLM outputs. We reframe hallucination detection as a point-wise anomaly detection problem by treating logits as time-series data. Our approach extracts features from token-level logits, addresses class imbalance with SMOTE, and trains an XGBOD model for probabilistic character-level predictions. Our system, which relies solely on information derived from the logits and token offsets (using pretrained tokenizers), achieves competitive intersection-over-union (IoU) and correlation scores on the validation and test set. 

A series of research projects focused on sentiment analysis in code-switched and African languages using multilingual embeddings, transformer models, and novel translation and model fusion techniques, contributing to multiple peer-reviewed publications and benchmark challenges. 

This project augments US Navy ship images by generating high-quality zoomed-out versions using OpenCV and Adobe Photoshop’s Generative Fill. Evaluated using a color-coded system, the study highlights AI’s potential for dataset enhancement in defense and autonomous systems while emphasizing rigorous evaluation and human oversight for reliability and trustworthiness. 

My research project aims to improve speech recognition for African American Vernacular English (AAVE) speakers by addressing bias in existing technologies. We've collected over 200 hours of AAVE audio to train new models and evaluate current systems, highlighting performance gaps and developing more accurate, inclusive solutions.

This project introduces a three-step GPT-based pipeline that enhances machine translation by pre-extracting and refining named entities using Wikidata. By isolating and accurately processing entity names, the system significantly improves translation fidelity, especially in complex or low-resource languages. 

The objective of this project is to design a usable and explainable language model that can assist in tactical decision-making across various domains, including military operations and emergency response. The model will be piloted with real humans to determine the best ways to display the AI output and allow the human to exhibit less stress during high-stakes decision-making

This project explores the use of the BART-large-MNLI model for zero-shot classification of food recall documents which predicts hazard and product categories from unstructured text, enabling automated food hazard detection. The approach addresses challenges like imbalanced data and ambiguity while emphasizing explainability, making it a robust solution for enhancing food safety monitoring systems. 

Emotion detection in text is hard due to cultural and contextual differences. The team used small expert models and a larger model to combine their insights. The system worked well for high-resource languages but struggled with low-resource ones. 

Developing a pipeline to process newspaper images from microfilm reel by page segmentation and metadata extraction using masking and segmentation models. The system will be able to crop pages out, detect the publication name, date of issue and page numbers and apply color presets and batch OCR on them following FADGI standard. The pdf of every single issue will be uploaded to cloud for final storage with searchability features. 

The project develops a networked co-management system that integrates real-time production control via a PLC (with pumps and valves) with a software-based workforce management system. This integration enables simultaneous management of production levels and workforce status, adapting operations based on product availability and staff capacity.