Research & Projects
Publications & selected work
Hover a card to flip it and read the details. On mobile, tap to flip. Full list on Google Scholar.
Benchmarking Causal Discovery Algorithms for Scale Free Graphs
Simulation study report (PDF)
Hover to learn more
- Benchmarked BOSS, FGES, PC, and DAGMA across Erdős–Rényi vs scale‑free DAGs (hub/sink structures).
- Varied network size/degree and sample size; compared precision/recall/F1 and calibration metrics.
- Key finding: BOSS was most robust on scale‑free topologies, maintaining high precision with stronger recall than alternatives.
Choose Better Hyperparameters
Simulation study report (PDF)
Hover to learn more
- Extensive simulation sweep to guide hyperparameter selection for score‑based causal discovery (BIC / penalty discount).
- Evaluated multiple graph configurations (ER vs scale‑free), sizes (20 vs 100 vars), degrees (2 vs 6), and sample sizes.
- Practical takeaway: penalty discounts around ~0.6–1.0 often gave the best F1 (balancing precision/recall), especially with resampling.
Resampling × Causal Discovery Simulation Study
arXiv preprint · With B. Andrews & E. Kummerfeld
Hover to learn more
- An extensive simulation study evaluating the interaction of resampling techniques across multiple causal discovery contexts.
- Systematically compares how resampling choices interact with causal discovery performance across diverse settings.
- Authors: Ritwick Banerjee, B. Andrews, E. Kummerfeld (2025).
Student Achievement Analytics
IHI · Dr. Erich Kummerfeld's group
Hover to learn more
- Investigated racial and economic achievement gaps across U.S. public school districts.
- Built descriptive models linking Panorama survey responses to academic outcomes.
- Developed predictive models for subject-level success using historical performance data.
- Managed pipelines for 150k+ students over 18+ years — cleaning through deployment.
Authorship Attribution
NLP class project
Hover to learn more
- Classified text authorship using n-gram features and a RoBERTa sequence classifier.
Multi-Class Cardiovascular Condition Classification
Machine learning project
Hover to learn more
- Multi-class classification pipeline for cardiovascular conditions.
Diagnosis of AD vs PSP
Neurodegenerative disease ML
Hover to learn more
- Exploratory ML analysis for differentiating Alzheimer's disease and PSP.
Talkative-LLM Evaluation Pipeline
Minnesota NLP Group (past) · UMN
Hover to learn more
- Designed an end-to-end pipeline to evaluate LLMs on MMLU, HellaSwag, Winogrande, ARC, GSM8K, and more.
- Supported API-based, Hugging Face, and custom model backends.
- Built cluster scripts for distributed evaluation and comparative performance visualization.
Healthcare Pre-Screening Chatbot
CS 5541 (NLP) · University of Minnesota
Hover to learn more
- LLM-powered chatbot to pre-screen patients based on symptoms and medical history.
- Suggested diagnoses, diagnostic tests, and specialist referrals.
- Used RAG and conversational memory via LangChain for personalized responses.
- Deployed with a Streamlit front-end and FastAPI back-end.
Disease Prediction with Decision Trees
Capstone · KIIT University
Hover to learn more
- Decision-tree classifier predicting diseases from symptoms at 92.7% accuracy.
- Integrated with a physician database for specialist referral.
- Optimized via pruning, Ant Colony Optimization, and BAT algorithms.