Projects

HALT: Hallucination Assessment via Log-probs as Time series

A lightweight black-box hallucination detector using token log-probabilities as a time series

Notes: arXiv preprint; introduces HUB benchmark spanning 10 LLM capabilities; 30× smaller and 60× faster than encoder-based baselines

Tech: LLMs, Time-series Classification, GRU, Calibration Modeling, vLLM, PyTorch

View Project

MuDoC: A Multimodal Document-grounded Conversational AI System for Classrooms

An interactive and trustworthy AI system for students to study course material

Notes: Presented at AAAI-MAKE 2025 and AIED 2025, increased student engagement with course materials, developed a trustworthy AI system for classrooms

Tech: LLMs, Interleaved Text-and-Image Generation, Multimodal Embedding Models, Document Layout Analysis, Weaviate VectorDB, AWS, ReactJS, Docker

View Project

Jill Watson: An LLM-based Virtual Teaching Assistant

A conversational AI teaching assistant that enhances teaching and social presence in online learning environments

Notes: Presented at AIED 2024 and Learning@Scale 2024, improved teaching presence in classrooms, supported thousands of students across multiple courses

Tech: OpenAI LLMs, Skill-based Architecture, Learning Analytics, VectorDBs, Learning Management Systems, Learning Tools Interoperability

View Project

MetaCues: Enhancing Critical Thinking in Generative AI Use with Metacognitive Cues

A system that enhances critical thinking in AI-powered search through metacognitive cues

Notes: Research project focused on improving user's critical thinking skills when interacting with generative AI search systems through strategic metacognitive interventions

Tech: LLMs, Metacognitive Cues using OpenAI API, Search Interface, ReactJS, Flask App

View Project

Protecting Human Cognition in the Age of AI

A study of frameworks to understand impact of AI on human cognition

Notes: Work presented at CHI 2025 Tools for Thought Workshop

Tech: Cognitive Science, Human-computer Interaction Design, Educational Technology

View Project

ALC3: Active Label Correction for Modular AI Systems

Novel method for improving LLM-based modular AI systems through iterative active label correction

Notes: Work presented at EMNLP 2024, our novel method ALC3 improved over existing baselines with 17-24% lower human feedback requirements

Tech: Large Language Models, Active Label Correction, Natual Language Processing, Modular AI Systems, HuggingFace Transformers, OpenAI GPT-3.5

View Project