Replicated key results from In-context Learning and Induction Heads (Olsson et al.) using TransformerLens on GPT-2 Small. Traced induction head formation through attention pattern visualization and targeted activation analysis. Wrote up the methodology and results, including where the replication held and where it needed adjustment.

AI workflow automation agent integrating LLM APIs with Slack and Google Workspace via LangGraph. Designed a stateful multi-step reasoning system handling NLP parsing, validation, storage, and API execution. Investigated failure modes in tool-use chains — where multi-step agents break down and what that implies for multi-agent safety.

Offline-first Next.js application for tracking community program delivery in low-connectivity environments. Designed a full-stack sync architecture with Supabase, local persistence, and batched sync — where the main challenge was maintaining data integrity across unreliable connections. Built for real deployment at the Conway Innovation and Entrepreneurship Center.

End-to-end ML pipeline on 864 audio recordings from 108 patients. Trained an ensemble of Random Forest, Gradient Boosting, and Hidden Markov Models, reaching F1 = 0.71 and precision = 0.80. Investigated feature attribution to understand what the models were actually using — an early exercise in interpretability in a high-stakes setting.