Large language models (LLMs) such as ChatGPT, Claude, and Llama-2 are extensively used for personal and business needs,
achieving widespread popularity and integration into numerous applications.
Despite their great success, LLMs tend to “hallucinate” and may generate inaccurate/fabricated information.
Due to the increased reliance on LLMs, the hallucination problem can pose a serious threat to security and degrade users’
trust. For example, the use of LLM-integrated applications in medicine may generate medical reports that contain highly
convincing yet entirely false information, possibly leading to life-threatening treatments or lack thereof.
Therefore, detecting hallucinations of LLMs is of primary importance..
The prediction of protein-protein interactions (PPIs) plays a vital role in understanding cellular processes,
disease mechanisms, and therapeutic target identification. Experimental methods for identifying interacting protein pairs,
are labor-intensive, time-consuming, and challenging to scale up due to the vast diversity of proteins.
Additionally, traditional computational methods using deep learn- ing to address this problem, although effective to some
extent, face challenges in terms of accuracy and efficiency.
This project proposes an improved approach by integrating science-guided machine learning (SGML) into PPI prediction models.
Two-Cloud Private Read Alignment to a Public Reference Genome (2020-2022)
The large-scale nature of human genome necessitates outsourcing of genomic computations to public clouds.
However, this raises some serious privacy concerns.
We designed and developed a two-cloud private read alignment algorithm using the Burrows-Wheeler Transform and the FM-Index. Our paper is
accepted to Proceedings of Privacy Enhancing Technologies Symposium 2023.
Graph Sketches and Embeddings: A Study of their Applications on Graph Databases (Master's Thesis) (2019)
An important practical consideration while processing large volumes of graph data is creating summaries of the data,
since it is usually impractical to perform operations on the entire dataset. In my thesis, I studied graph sparsification and embedding
techniques and provided a performance study of thier effects on graph properties like page rank, betweenness centrality, community detection
and node similarity.
Evaluation of High-Performance Application-level Caching of Generic Graph Structured Data (2018)
In addressing the challenge of handling large graph data, my project focused on hashing techniques and their impact on graph processing.
I studied various hashing methods and implemented foundational graph algorithms like shortest path, and minimum number of hops, using SNAP datasets.
I evaluate these results in the neo4j graph database, where I assessed the influence of these techniques on the efficiency and accuracy of graph-related operations.
This work contributes to understanding how different hashing approaches can optimize graph data processing.
Supporting Non-relational Models with Flexible Schema Storage (2017)
NoSQL databases offer a unique advantage over traditional relational databases, particularly when handling complex and interconnected data.
This project involves comparing the performance of NoSQL data stores with that of relational data models. Additionally, I explored tile-based
architectures and provided a theoretical discussion on what the mappings from various NoSQL data stores and the relational model to tile-based
architectures might look like.
