Projects

Research and engineering work in ML systems, LLM trustability, and explainable AI

Research LLM Security

Automated LLM Penetration Testing

Benchmarking and improving LLM security through automated vulnerability assessment.

Overview

Developed a comprehensive benchmark for evaluating LLM security vulnerabilities, including prompt injection, jailbreaking, and data extraction attacks. Created automated testing pipelines that identify weaknesses in production LLM deployments.

Key Contributions

  • Novel benchmark suite for LLM security evaluation
  • Automated attack generation and response analysis
  • Defense mechanism recommendations based on vulnerability patterns

Tech Stack

Python, LLMs, Security Testing Frameworks, Prompt Engineering

Related Publication →
Research Privacy

Secure Multiparty Generative AI

Enabling collaborative AI without exposing sensitive training data.

Overview

Research on secure computation methods for generative AI, allowing multiple parties to jointly train and query models without revealing their private data. Addresses key challenges in federated learning and secure inference.

Key Contributions

  • Privacy-preserving inference protocols for LLMs
  • Secure aggregation methods for multiparty training
  • Threat model analysis for collaborative AI systems

Tech Stack

Cryptographic Protocols, Federated Learning, PyTorch, Secure Computation

AAAI 2025 Workshop →
Research XAI

Knowledge-Based Explainable AI

Making AI explanations meaningful through case-based reasoning.

Overview

Research on explanation systems that go beyond feature attribution to provide human-understandable rationales. Combines case-based reasoning with modern ML to generate explanations grounded in domain knowledge.

Key Contributions

  • Integration of CBR with neural network explanations
  • Explanation containers for biomedical QA systems
  • User studies on explanation effectiveness

Tech Stack

Case-Based Reasoning, Knowledge Graphs, Python, NLP

Research Healthcare

Medical Knowledge Extraction with LLMs

Structured extraction from clinical text for knowledge graph construction.

Overview

Developed LLM-based pipelines for extracting structured medical knowledge from unstructured clinical text. Enables automated construction of medical knowledge graphs for downstream search and summarization tasks.

Key Contributions

  • Entity and relation extraction from clinical notes
  • Knowledge graph construction from extracted triples
  • Integration with medical search and summarization systems

Tech Stack

LLMs, NER, Relation Extraction, Neo4j, UMLS

Research Computer Vision

Synthetic Image Detection (E3)

Detecting AI-generated images with limited training data.

Overview

Ensemble approach for detecting synthetic images from new generators using limited training samples. Addresses the challenge of rapidly evolving image generation models and the need for adaptable detection systems.

Key Contributions

  • Expert ensemble architecture for few-shot adaptation
  • Robust detection across multiple generator architectures
  • Efficient fine-tuning strategies for new generators

Tech Stack

PyTorch, Vision Transformers, Few-Shot Learning, Image Forensics

CVPR 2024 Workshop →
Engineering Data

Enterprise Data Backbone

Scalable data infrastructure powering analytics and ML workflows.

Overview

Designed and deployed a multi-organization datastore serving as the backbone for analytics and ML pipelines. Built for 24/7 reliability with automated ETL, real-time transformations, and comprehensive monitoring.

Key Features

  • Automated ETL pipelines with Airflow orchestration
  • Real-time streaming transformations with Kafka
  • Data quality monitoring and alerting
  • Self-service analytics layer for business users

Tech Stack

Apache Airflow, Spark, Kafka, AWS, PostgreSQL, dbt

Upcoming Finance

Financial Trading ML Analysis

Trustworthy AI for financial decision making.

Overview

Starting 2026: Research on applying trustworthy AI principles to financial trading datasets. Focus on interpretable models, uncertainty quantification, and knowledge-grounded predictions for high-stakes environments.

Planned Focus Areas

  • Time-series modeling with uncertainty quantification
  • Knowledge-grounded market signal extraction
  • Interpretable trading strategy analysis

Tech Stack

LLMs, Time-Series Models, Knowledge Graphs, Risk Analysis