PhD Researcher · Statistics & AI
Nischal
Subedi
University of Delaware · Applied Economics & Statistics
I build statistical methods for reliable machine learning, focusing on understanding when models adapt in the wrong direction, retrieve the wrong evidence, or are confidently wrong. My work is motivated by theory (random matrix theory, geometric probability) but validated empirically on real NLP benchmarks. More broadly, I am interested in high-dimensional inference, representation learning, and uncertainty quantification as foundations for trustworthy AI systems.
Latest
News
Apr 2026
AgreRank accepted to SIGIR 2026 Full Papers Track (18.4% acceptance rate). Code on GitHub →
2026
NeurIPS submission in preparation — "When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation." Introducing PivotLoRA and a regime boundary at d/n ≈ 10.
Nov 2025
Article published in ACM Interactions Magazine — "Empowering Tenants and Landlords Through Conversational AI for Housing Rights."
Aug 2025
Paper published in Urban Findings — "A Proactive Framework for Identifying At-Risk Affordable Housing in High-Pressure Markets."
Jun 2025
Featured in Shelterforce Magazine — Landlord-Tenant Rights AI Assistant highlighted as a tool for tenant advocacy.
Aug 2025
Started PhD in Statistics & Data Science at University of Delaware under Dr. Cencheng Shen. Enrolled as Research Assistant.
Feb 2025
Earned Graduate Statistician (GStat) designation from the American Statistical Association.
Work
Research
NeurIPS 2026 (in prep) · PEFT / LoRA
When Does Subspace Direction Matter for LoRA? Regime Analysis of the Magnitude Principle in Few-Shot Adaptation
We challenge the dominant "magnitude principle" in LoRA initialization and show that subspace direction — not magnitude — governs adaptation quality when d/n > 10. We introduce a regime boundary framework supported by Random Matrix Theory, an ablation ladder (EigenLoRA → LDA-LoRA → PivotLoRA), and empirical validation across 30 BERT settings and Llama-3.1-8B.
PivotLoRA
Few-Shot
Random Matrix Theory
BERT
Llama-3.1-8B
PEFT
SIGIR 2026 · Full Papers · Melbourne · Acceptance Rate 18.4%
AgreRank: Consensus-Anchored Expansion for Noise-Resilient Re-ranking
AgreRank fuses sparse (BM25) and dense retrieval signals via percentile normalization and geometric gating to form a consensus seed set. This reduces semantic drift and improves nDCG@10 by up to 5.7% over RRF+CE baselines on TREC Deep Learning benchmarks. We evaluate across BGE, E5, and multiple reranker architectures.
Data & Policy (under review) · Graph ML / Housing
A Graph Neural Network Approach to Physical Distress Screening in Subsidized Housing
We apply Graph Attention Networks to predict physical distress in subsidized housing, incorporating urbanicity stratification and property-level graph structure. This work supports proactive policy interventions in affordable housing preservation.
Graph Attention Networks
Housing Policy
GNN
Social Impact
Output
Publications & Features
SIGIR 2026 · Full Papers
AgreRank: Consensus-Anchored Expansion for Noise-Resilient Re-ranking
Urban Findings 2025
A Proactive Framework for Identifying At-Risk Affordable Housing in High-Pressure Markets
ACM Interactions 2025
Empowering Tenants and Landlords Through Conversational AI for Housing Rights
Urban Findings 2025
Preserving the Subsidized Housing Safety Net: A Multi-Dimensional Risk Assessment Strategy
Shelterforce 2025
Tech Tools Help Tenants Push Back Against Problematic Landlords
NC J. Math & Stats 2019
Implementing the Lifetime Performance Index with a Two-Parameter Rayleigh Distribution Under Progressive Type II Censoring
MAA MathFest 2018
Harmonious Labeling and Graphs
Career
Experience
University of Delaware
Research Assistant
- Building AgreRank, a consensus-based hybrid retrieval framework; evaluated on TREC Deep Learning benchmarks.
- Designing and benchmarking LoRA variants (EigenLoRA, LDA-LoRA, PivotLoRA) across GLUE and SQuAD under few-shot conditions.
- Studying the statistical geometry of high-dimensional LLM embedding spaces using Random Matrix Theory.
Open Avenues
Data Science Fellow
- Directed a course on housing price prediction using XGBoost, covering EDA, feature engineering, and AWS SageMaker deployment with SHAP interpretation.
Home Partners of America
Data Scientist
- Built a dynamic pricing system using survival models, increasing leasing rates by 20% and reducing review time by 50%.
- Implemented a sentiment classifier with map visualization on AWS QuickSight, boosting CSAT scores by 35%.
- Developed a market monitoring system using web scrapers and MLS data to inform regional leasing strategies.
Barclays Bank
Risk Analyst
- Maintained a logistic regression model for high-risk account prediction; monitored data drift for compliance-aligned accuracy.
- Automated Python reporting scripts, reducing manual effort by 40%.
Auburn University
Summer Research Intern
- Developed harmonious labeling algorithms for planar graphs (even/odd paths). Presented results at MAA MathFest 2018.
University of Delaware
Graduate Teaching Assistant
- Assisted in teaching STAT 200: Applied Statistical Methods; tutored students, prepared exams, and generated score reports.
Community
Service & Outreach
Peer Review
Reviewer — Journal of Intelligent Manufacturing and Systems – Design (JIMS-D)
Reviewer — Data Science Journal (DSJ)
Reviewed 20+ manuscripts across ML, deep learning, predictive modeling, and applied AI.
Volunteer
DemocracyLab
Volunteer Tech Contributor
Contributing data science expertise to civic tech projects connecting volunteers with social good initiatives.
DataKind
Volunteer Data Scientist
Applying AI and ML to social impact projects for nonprofits; focused on predictive modeling and cloud deployment.
Toolkit
Technical Skills
Languages
ML / DL
LLMs & NLP
Cloud & DevOps
Search & Data
Statistics
Recognition
Awards & Certifications
Graduate Assistantship
University of Delaware, 2019
Chancellor's Award — Full Tuition Scholarship
Troy University, 2015–2019
AWS Certified Machine Learning
Dec 2024
Deep Learning Specialization
Coursera / DeepLearning.AI, June 2025
SAS Certified Statistical Business Analyst
Regression & Modeling, 2021
Society of Actuaries — Exam P
Sep 2018
Pi Mu Epsilon Honor Society
Inducted 2018
Chancellor's List (6×) · Dean's List (1×)
Troy University