An end-to-end scientific claim verification system built on a two-stage fine-tuned DeBERTa-v3-base with retrieval augmentation and explainability. Identified and corrected bilateral data leakage affecting 38.4% of the SciFact development split. Macro-F1: 0.9043 (MultiNLI), 0.8640 (SciFact, oracle), 0.7713 (SciTail, zero-shot). Deployed with FastAPI, Docker, and Vercel; integrates Semantic Scholar retrieval, Integrated Gradients (Captum), and GPT-OSS-20B explanations.

With S. T. Hmuu, School of Biosystems and Biomedical Sciences, Korea University.

A pre-registered, causal faithfulness study comparing sequence (ChemBERTa) and graph (GIN) neural network representations for interpretable molecular toxicity prediction, using perturbation-based metrics (comprehensiveness, sufficiency, deletion and insertion AUC) validated directly against curated ground-truth toxicophore substructures. Diagnosed a class-direction confound in metric aggregation that, left uncorrected, would have reversed one of the four primary findings, then resolved it through subgroup analysis and recomputed the full result set before drawing conclusions. Found ChemBERTa's attributions substantially and consistently more faithful than GIN's across all four metrics (n = 1,457, p < 10⁻⁸⁰), with the effect stable under molecular distribution shift and replicated on an independent, differently imbalanced endpoint (Tox21/SR-ARE).

An autonomous agentic system for citation-faithfulness verification, orchestrating dual independent verification (fine-tuned DeBERTa-v3 NLI + zero-shot LLM) with a LangGraph ReAct planner that resolves disagreement via tool use — no human intervention required. Caught a shared-bias case where both verifiers agreed on an incorrect SUPPORT verdict at 0.9931 confidence, and corrected it to NOT_ENOUGH_INFO through autonomous escalation across three bounded agent actions, streamed live via Server-Sent Events.

Diagnosed representation anisotropy in transformer encoder outputs — tracing the cause through recent literature (Godey et al., 2024) — and corrected it via GloVe-based embedding initialization, validated through PCA geometric analysis. Built the full encoder from scratch in NumPy (multi-head self-attention, sinusoidal positional encoding, GELU, layer norm, residual connections, following Vaswani et al. 2017) to get direct inspection and control over the representation space itself.