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.

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.

A full transformer encoder implemented from scratch in NumPy — multi-head self-attention with padding masks, sinusoidal positional encoding, GELU activation, layer norm, and residual connections, following Vaswani et al. (2017). Diagnosed anisotropy in randomly initialized encoder outputs, traced its cause through the literature (Godey et al., 2024), and applied a GloVe-based mitigation with PCA geometric analysis to produce interpretable contextual representations.