I build systems that turn messy data into decisions.

Problem Life-insurance underwriters read multi-page application packets by hand. The brief: an AI copilot that screens financial viability and routes the easy cases automatically.

Built A scoring pipeline — six-check conflict screen, a weighted rule engine with mortality-evidence-anchored weights (NHANES + NCHS linked mortality data), logistic regression and gradient-boosting models, and an affordability screen — plus a single-file workbench UI that re-derives every score in the browser from exported model coefficients.

Problem Reward models used to score people — loan applicants, job candidates — inherit bias from their training data.

Built A curiosity-driven reward debiasing system; I ran the experiments on job-market and lending datasets, debugged training runs, and documented the evaluation methodology.

Problem Clinicians comparing meningitis case reports have no fast way to find similar past cases in unstructured clinical text.

Built An NLP retrieval pipeline matching case reports with OpenAI embeddings + FAISS vector search, with an interactive interface for case exploration and automatic summaries.

Problem Real-time verification of biological samples for the Harvard HSRC Innovation Challenge — and it had to survive a live on-stage demo.

Built The full prototype: Flask backend APIs, detection logic, a React dashboard, Docker packaging, and latency tuning for the stage.

Problem Turn raw interaction data — keystrokes, click patterns — into real-time cognitive-performance insights people can actually read.

Built Backend features and Python data pipelines supporting secure time-series tracking, working directly with the founders on priorities and product refinement.