• Formulated inference-time bias correction as an online change-point detection problem: a per-step bias-risk signal feeds a CUSUM statistic, and a targeted correction fires only once accumulated evidence crosses a calibrated threshold — rather than at fixed intervals.
• On gpt-4o-mini, adaptive triggering recovered most of the disambiguated-context accuracy that fixed-interval intervention destroys (90.1% vs. 82.9%, against a 92.1% no-intervention baseline) using 0.60 rather than 1.00 interventions per item. The advantage holds with an independent judge (+6.8pp, 95% CI [+2.8, +10.7], p = 0.0015).
• Compared white-box (next-token logits) and black-box (LLM-judge) bias-risk signals across seven open-weight models and GPT-4o-mini, running large-scale experiments on ASU’s Sol supercomputer. The white-box signal improved ambiguous-context accuracy on all six evaluated models but hurt disambiguated accuracy on five — it cannot separate unsupported stereotype reliance from correct, stereotype-congruent evidence. Adaptive timing cannot rescue a misaligned signal.
• Found and corrected four undocumented schema errors in the BBQ benchmark that had left 17–50% of items in affected categories mis-scored, and open-sourced the fix.

Read the paper Code

• Showed in a 2³ factorial human-subject study (n = 304) that high-transparency LLM outputs did not improve performance and in marginal conditions decreased it — evidence of overreliance from information overload, motivating adaptive / on-demand transparency design.
• Built MASTOPIA, a multi-agent RAG system (supervisor → retriever → generator agents) powered by GPT-4 / GPT-3.5 that operationalizes Multisource AI Scorecard Table (MAST) tradecraft standards through prompt engineering.
• Shipped an interactive Streamlit demo with model conditions, evidence retrieval, and behavioral logging of verification activity.

Try the demo Code

• Mitigated social biases in T5, BERT, and LLaMA-2 for toxicity and hate-speech detection by combining reinforcement learning with parameter-efficient tuning (LoRA, P-tuning).

• Ran human-subject studies on whether off-the-shelf or improved GPT-4 outputs lead to appropriate use — including correct rejections — for intelligence-analysis (I&A) tasks.
• Applied the MAST trust-assessment framework to evaluate baseline performance and inform adoption decisions for GPT-4 in I&A workflows.
• Improved response quality with prompt engineering and RAG across summarization, NER, and conversational tasks.

• Trained an LLM classifier to score document–summary pairs via multi-class classification + RLHF on a handcrafted human-preferences dataset; validated with expert evaluation to confirm the learning method.

• Designed and validated a principled AI design framework (PADTHAI-MM) for trustworthy decision-support systems; demonstrated effectiveness through a deployed AI system that positively impacted user trust perceptions.
• Conducted association analysis between user ratings and trust-impacting factors, providing a theoretical basis for the framework.
• Released open-source implementation artifacts for the READIT and Facewise prototypes supporting the MAST-based design workflow.

Code

• Built a Transformer-based summarization system for intelligence analysts, deployed via a Node.js + Google Cloud web interface for production access to summarized reports.
• Public implementation available through the PADTHAI-MM research-code repository.

Code

• Built a face ID verification system for security screening, using CNN + ResNet face matching with fine-tuning to optimize verification performance.
• Public implementation available through the PADTHAI-MM research-code repository.

Code

• Final project for CSE 472 (Social Media Mining).
• Implemented counterfactual explanations for a multi-layer neural network used in text classification.

Project Report

• Extracted biomedical entities and identified relations using the Comparative Toxicogenomics Database (CTD) via distant supervision.
• Implemented and trained a tree-RNN model (SPINN) combined with a word–character embedding model.

MASTOPIA is a Streamlit demo for LLM-assisted intelligence analysis. It operationalizes MAST-style transparency ideas through an interactive workflow with model conditions, evidence retrieval, model information, session management, and activity logging.

Try the demo View code

System highlights

• Multi-agent architecture with supervisor, retriever, and general AI agents.
• Streamlit chat interface with model information controls and sidebar state.
• Session management and activity logging for research evaluation.
• RAG-oriented stack using LangChain / LangGraph, OpenAI APIs, FAISS, and Google Cloud / Firestore components.

Sprout is not a research project, but it shows my interest in building usable real-world systems with care for interaction quality, reliability, and user-facing product details.

Try Sprout View code

• Full-stack bilingual social app with authentication, profiles, public sharing, responsive UI, dark mode, and Supabase-backed persistence.
• Built with a high-quality agentic coding workflow while retaining ownership of product direction, architecture, testing, and final code quality.
• Included here as engineering evidence alongside my primary research work.