Stress-Test RAG in Biomedical QA
End-to-end pipeline combining dense retrieval, re-ranking, and LLaMA-3-based answer generation for BioASQ.
Hi, I’m Tanushree 👋
AI Researcher & ML Engineer. I got curious about the difference between a model performing well and a model actually being trustworthy turns out that gap shows up everywhere, in biomedical AI, in trading systems, in anything making decisions under uncertainty.
About Me
I study where AI systems fail not what they achieve under ideal conditions, but how they break down under noise, conflicting evidence, or incomplete information. My research spans biomedical RAG reliability and applied security, with three accepted papers so far, alongside deep reinforcement learning for decision-making under uncertainty. The throughline is one question: how do we know when a model can be trusted, and when it shouldn't be. Open to PhD research and applied ML/data science roles.
Featured Projects
Kalshi Algo Trading & Arbitrage
High-frequency trading system in Rust with Deep Reinforcement Learning (DQN) in PyTorch to execute short-term arbitrage.
Private deployment — live dashboard not publicly hosted; walkthrough available on request.
Agentic Web-Retrieval Engine
Autonomous research agent performing multi-step workflows using instruction-tuned LLMs and web retrieval.
Ongoing Research
Agentic AI for Cardiac Decision Support
Building an agentic clinical decision-support system for cardiac care. Developed a two-branch 1D CNN + RR-interval arrhythmia classifier over 109K beats (AAMI 5-class), with controlled DS1/DS2 ablations evaluating macro-F1, average precision, Brier score, and ECE for downstream agent reliability.
Fairness vs. Accuracy in Machine Learning Models
Built a PyTorch training method with a tunable accuracy–fairness trade-off, evaluated across five real-world decision-making datasets (criminal risk, income, credit default, health coverage). Reduced demographic-group disparities by up to 93% at under 0.5% accuracy loss.
Publications
Stress-Testing Biomedical Retrieval-Augmented Generation
Nepal, T. et al. “Stress-Testing Biomedical Retrieval-Augmented Generation: A Diagnostic Evaluation under Noise, Conflicting Evidence, and Unanswerability.” IEEE HealthCom 2026, New York, NY, Oct 19–21, 2026.
CodeBERT at the Watering Hole
Dadala, K., Nepal, T., Tsai, I.H. “CodeBERT at the Watering Hole for Dual-Layer URL and JavaScript Attack Detection.” IEEE EIT 2026. Presented May 22, 2026.
What Hurts RAG in Biomedical QA
Nepal, T., Tsai, I.H. “What Hurts RAG in Biomedical QA: A Stress-Test Study of Noise, Conflict, and Unanswerability.” MMLS 2026, Purdue University, June 25, 2026.
Generative Semantic Enhancement for Short-Video Recommendation
Nepal, T. et al. “Generative Semantic Enhancement for Short-Video Feed Recommendation: A Converged Study on KuaiRec.” IEEE Big Data 2026, Phoenix, AZ, Dec 14–17, 2026.
Online Cyberbullying Detection Using NLP
Nepal, T. (2026). “Online Cyberbullying Detection Using Natural Language Processing.” Zenodo. SVM-based detection over a 24k+ tweet dataset.
GenAI Impact on Learning
Quantitative study applying correlation, regression, and mediation analysis on generative AI influence on students.
Skills
Loading skills…
Experience
Loading experience timeline…
Let's Connect
Feel free to reach out for collaborations or just a friendly hello. Always open to discussing Data Science and AI.
Live System Traffic
Real-time Map Monitoring
Total Visits
--
--
Initializing connection...
Competency Map
Education
Loading education timeline…
Blog / Articles
Gross Domestic Product (GDP) — Analysis
Comparing developed vs developing nations, key indicators, and a simple linear regression model.
Photography
Sky & landscape shots I've taken.
Photos coming soon — drop images into assets/photography/ and they appear here.
Resume / CV
Three versions, depending on what you're reading for. Each opens as a PDF.