When can models and data be trusted?

With Dr. Christine DeLorenzo

• Onboarding to the actigraphy quality-control workflow for health research data.

With Prof. Rahmati and Kevin Eykholt

• Question: do search rankings stay stable when the same question is reworded without changing its meaning?
• Method: a reproducible pipeline with cached query results comparing Google and DuckDuckGo on the same reworded queries.
• Defined True-PIR, a demotion-only ranking-stability score, plus bias diagnostics.
• Result: True-PIR ≈ 0.625 for Google vs ≈ 0.586 for DuckDuckGo (subset). Written up as a Medium article, not a paper.

• CNN-based malaria detection from blood smear images (92% accuracy).
• Led the research track for a Smart India Hackathon national finalist project (Ministry of Power problem statement).

Aegis: cost-aware fraud detection under drift

Swept 19 decision thresholds on synthetic PaySim data (49,962 transactions, 0.138% fraud; 13 fraud cases in the held-out set) with assumed costs of $1.5M per missed fraud and $15 per false alarm. Cost-optimal threshold 0.20 ($40,140); operational threshold 0.70; $19.5M modelled cost gap; weighted F1 0.994. Drift detection validated with controlled shifts: PSI 0.003 → 0.640, KS 0.014 → 0.304. Synthetic data, so results are an upper bound.

Negation and uncertainty in rare disease notes

NLP pipeline that separates present, negated and uncertain findings in clinical notes and maps them to HPO and GA4GH. Finalist, Harvard HackRare 2026 (team of 3).