Hi, I'm Prayash
/pray-yash joe-shee/
I am interested in systems designed to make Large Language Models verifiable and trustworthy. I hold a Bachelor's degree in Computer Science with a minor in Statistics and recently completed graduate coursework at Virginia Tech.
My research experience spans building policies for evaluating sustainability , formal test verification for blockchain contract security , and graph-based reasoning extraction . For my thesis , I investigated whether LLMs can generate robust test suites by evaluating against code mutation.
Research Interests:
- Human-AI Teaming in Medicine & Education
- Formal verification for neural code generation
- Robust Generative UIs
- Metacognition in High-Stakes decision making
Research Interests:
- Human-AI Teaming in Medicine & Education
- Formal verification for neural code generation
- Robust Generative UIs
- Metacognition in High-Stakes decision making
I am exploring PhD opportunities focused on interpretable AI systems, verifiable generative interfaces, and human-centered AI for education.
Check out some of the potentials of Generative UIs in Learning and Training Settings
Past projects
See all projects-
Evaluating the Robustness of LLM-Generated Tests via Mutation AnalysisMaster's thesis investigating whether LLMs can generate test suites robust enough to detect bugs using mutation testing as ground truth. -
LLM-Enhanced Smart Contract Fuzzing for Ethereum Pectra SecurityAddressing security risks in the Ethereum Pectra upgrade by integrating LLMs with mutation-based fuzzing to detect vulnerabilities. -
High-Performance Code Generation Using LLMs: Tensor ReductionInvestigation of LLM capabilities in automating high-performance CUDA kernel generation for tensor reduction on NVIDIA A100 GPUs. -
From Context to Deception: Simulating and Detecting LLM-Driven Impersonation AttacksExploring the feasibility of automated voice phishing and developing detection mechanisms for LLM-driven impersonation attacks. - DebGraph: Graph Based Debate Evaluation and FeedbackA system combining Knowledge Graphs with LLMs to overcome limitations of subjective debate evaluation and provide accurate winner predictions.
Past Presentations
See all presentations
Latest posts
See all postsNothing yet! Check back later :)