My research asks how a prediction should carry honest uncertainty — and what it takes to earn it.
A weighted ensemble of ARIMA, ETS, regression and tree-based learners predicting IBM's global revenue across business units — delivering intervals, not just point estimates. Built on statsmodels, pmdarima, statsforecast and scikit-learn, with ensemble weights optimised per unit against held-out validation.
$62B revenue under forecastMy dissertation: a Matrix Ensemble Kalman Filter multi-arm neural network that combines several deep learners in situ — matching transfer-learning accuracy while producing tighter prediction intervals at higher coverage than Monte Carlo Dropout. Applied to cancer drug-response prediction with Argonne National Laboratory.
PhD, Statistics · Nebraska 2023Gut bacteria digest complex sugars using clusters of genes that act together. Which sugar a given cluster targets normally has to be established in the lab. Treating a cluster as a sentence of gene “words” lets representations be learned from abundant unlabelled sequence, so the few laboratory-confirmed examples are spent on the classifier rather than on learning the vocabulary.
Nucleic Acids Research · Sensors · GeneticsAn enterprise RAG assistant built end to end — data scoping, web scraping, ingestion into Milvus and FAISS, and a benchmark of open-source against proprietary foundation models on retrieval quality, latency and cost-performance.
8,000+ IBM employees servedA multi-agent orchestrator for IBM's marketing teams. It ingests raw event material — prospectuses, briefs, supporting assets — and coordinates tool calls across a multi-turn pipeline of upload, multi-format extraction, embedding and retrieval-augmented generation, so the copy it produces stays grounded in the source documents rather than invented.
Deployed for IBM marketing teamsNearly three years on staff at IBM's Chief Analytics Office, preceded by research internships at LinkedIn, Seagate, Travelers and Hennepin County. Each was a chance to test statistical methodology against a problem someone actually needed solved, on a deadline, with real data.
IBM · Chief Analytics Office
Statistical and machine-learning methodology applied across the company, from classical time-series forecasting through to large language models and multi-agent systems. Four systems reached production:
IBM · Chief Analytics Office
Built machine-learning models recommending strategic products to IBM clients, paired with an interpretability framework that explained each recommendation in terms of competitive parity — so a seller could say not just what to offer but why it beat the alternative.
LinkedIn · Premium Data Science Group
Two pieces of work in the Premium subscriptions group:
Seagate Technology · Global Wafer Systems
Deployed a recommendation system that surfaced internal analytics reports to employees across two global sites, and built the surrounding machinery to keep it honest:
Hennepin County · Facility Services
Built statistical models quantifying the reduction in expenditure attributable to energy-saving practices in county-owned buildings — producing evidence-based justification for the money spent on them — and presented the findings to the energy managers of Hennepin County and the City of Minneapolis.
Travelers Insurance · Personal Insurance R&D
Developed models predicting the occurrence and severity of fire claims on home insurance policies, and engineered a substitute for an expensive third-party data attribute using inexpensive data already in hand — which gave the company renegotiation leverage with the vendor. Presented the methodology and findings to both technical and non-technical stakeholders.
Five continuous years of graduate appointments at Nebraska, spanning research and instruction, preceded by three teaching assistantships at Minnesota.
University of Nebraska–Lincoln & Argonne National Laboratory
Two research lines, running in parallel:
University of Nebraska–Lincoln · Department of Statistics
Held continuously across five years, covering successive teaching, instructional and research roles:
University of Minnesota–Twin Cities · School of Statistics
Three teaching assistantships covering laboratory and discussion sessions, instructional support, grading of assignments and examinations, and office hours — for Statistical Analysis, Applied Statistics I, and Sampling Methodology in Finite Populations — reaching roughly 90 graduate and undergraduate students.
Peer-reviewed journal articles
Genetics 229(2), iyae195 · 32 citations
Sensors 24(7), 2172 · 14 citations
Nucleic Acids Research 51(D1), D557–D563 · 99 citations
Doctoral dissertation
PhD thesis, University of Nebraska–Lincoln
Preprints & manuscripts in preparation
arXiv:2307.10436 · journal version in preparation
In preparation
Book chapter
Springer · Festschrift in honour of Prof. Tathagata Bandyopadhyay
Conference proceedings
ESS Open Archive · North American Plant Phenotyping Conference
Further refereed presentations at WNAR 2023, the International Indian Statistical Association Conference 2023 and ISMB 2022 are listed on Google Scholar.
Four refereed presentations at the field's recognised meetings — two statistics societies, the flagship computational-biology conference, and the plant-phenotyping network.
V Piyush, G Fernando, S Ghosh
Presented the in-situ ensemble approximation: shallow networks trained alongside a deep multi-arm learner can stand in for it while carrying calibrated uncertainty — the practical core of the stacking method.
V Piyush, Y Yan, Y Zhou, Y Yin, S Ghosh
The methodological talk on the dissertation itself — treating stacking weights as an unobserved state updated by an ensemble Kalman filter, and what that buys in interval coverage and width.
V Piyush, R Zhang, R Pal, S Ghosh
Introduced a predictability score over protein sequences and showed it improves downstream classification. ISMB is the flagship annual meeting of the International Society for Computational Biology.
G Fernando, V Piyush, S Ghosh
Argued that field images are not independent observations but unequally-spaced sequences carrying information about the growth trajectory — and built a hybrid model that exploits that structure for tassel counting. Archived on the ESS Open Archive.
Each thread is backed by a public repository, with notebook outputs committed so figures and result tables read without running anything.
I have taught statistics from the first course a student ever takes through to graduate inference and experimental design — as sole instructor of record for three terms, and as a teaching assistant and laboratory instructor across eight more.
The practical application of statistical thinking to contemporary issues — collection and organisation of data, probability distributions, statistical inference, estimation and hypothesis testing. I was sole instructor across three terms, with full responsibility for the syllabus, delivery, assessment and grading — including taking the course fully online mid-pandemic.
A graduate course on the statistical characteristics of a research study — both designed experiments and studies where controlled experimentation is not feasible and its features must be mimicked. Covers power and the efficiency of competing designs, and the major design structures: blocking, nesting, multilevel models, split-plot and repeated measures.
An introductory course for researchers who will not go beyond STAT 802/803/804 — statistical concepts and methodology for the descriptive, experimental and analytical study of biological and other natural phenomena, weighted toward practical application rather than theory.
The theory course: interval and point estimation, sufficiency and completeness, Bayesian procedures, uniformly most powerful tests, the sequential probability ratio test, likelihood-ratio and goodness-of-fit tests, and elements of analysis of variance and nonparametrics.
Three invited guest sessions covering deep-learning foundations, convolutional networks on benchmark datasets, and pretrained transformer models — bringing modern machine learning into the mathematical statistics sequence.
Finite-population sampling: simple random, stratified, cluster, unequal-probability and systematic designs, with ratio estimators, regression estimators and model-based estimation.
The first semester of the applied sequence for statistics majors — a broad survey of applied methods, how to recognise which kind of problem you have, how to choose a method to match it, and how to read the result correctly, taught on real data throughout.
A fast-paced four-credit course for graduate students who need statistics as a research technique: descriptive statistics, elementary probability, estimation, one- and two-sample tests, contingency tables, correlation, linear and multiple regression, and analysis of variance — all on real data, in software.
Three days for graduate students and postdocs in the UNL Department of Food Science — an audience with deep microbiology expertise and no deep learning. Built backwards from a question they already cared about (what substrate does this gene cluster act on) rather than forwards from architectures, so each method arrived only when it was needed.
Why a gene cluster can be treated as a sentence, and what that buys you. Tokenisation choices and the consequences of getting k wrong.
Word2Vec, Doc2Vec and FastText trained on a large unsupervised corpus — so that representation learning never spends the scarce labelled examples.
Recurrent architectures trained by transfer learning on top of those embeddings, and reading attention weights as evidence rather than decoration.
Uncertainty Quantification for Deep Learning — calibration, conformal prediction, deep ensembles and Kalman-filter approaches, taught through healthcare and decision-support cases. Applied Generative AI: Retrieval, Evaluation and Governance — foundation-model selection, retrieval-augmented generation, and the statistical problem of evaluating systems whose output is text.
Three invited sessions in STAT 302, Mathematical Statistics and Modeling II covering deep-learning foundations, convolutional networks on benchmark datasets and pretrained transformer models — bringing modern machine learning into the mathematical statistics sequence.
Volunteer teacher of mathematics and statistics with Sweccha, a non-profit in India (2014–2015), working with children from underprivileged backgrounds and preparing lessons for widely varying levels of prior schooling.
Invited to speak at a national laboratory, at other universities' seminars and to domain departments outside statistics — where the job is to make a method land for an audience that does not share your vocabulary. The three-day workshop I designed is described under Teaching.
Invited talks & seminars
Master's Seminar, Department of Statistics, California Polytechnic State University
Argonne National Laboratory
with G Fernando, R Zhang, R Pal and S Ghosh
Department of Food Science, University of Nebraska–Lincoln
STAT 885, Statistical Data Mining and Machine Learning, University of Nebraska–Lincoln
Graphics Group, Iowa State University; and Department of Statistics Seminar, University of Nebraska–Lincoln
Delivered at two institutions
Department of Statistics Seminar, University of Nebraska–Lincoln
Department of Statistics Seminar, University of Nebraska–Lincoln; and Syngenta Statistics and Data Science Student Seminar
Delivered at a university and in industry
Poster presentations
Department of Statistics, University of Nebraska–Lincoln
Department of Statistics, University of Nebraska–Lincoln
Consulting is where a statistician earns their keep in a university — translating a collaborator's question into a design, and their data into a defensible answer. Mine has run from digital humanities through to enterprise deployment.
University of Minnesota Libraries & Zooniverse
Statistical and natural-language-processing consultant on a crowd-sourced transcription of U.S. Supreme Court Justices' handwritten conference notes. I advised on and designed the consensus-aggregation methodology — how to combine many volunteer transcriptions of the same manuscript line into one defensible reading — and explained the algorithm and its edge-case behaviour to political scientists and library staff.
University of Nebraska–Lincoln
Built on formal statistical-consulting training in STAT 8801 and STAT 825, and deepened by instructing STAT 802, Design and Analysis of Research Studies — the course that teaches researchers how to design a study that can actually answer their question, including when a controlled experiment is not available.
IBM Chief Analytics Office
Three years of translating research into production tools used internally and by business partners — the applied end of the same skill: understanding a stakeholder's real question, choosing a method that fits it, and being honest about what the answer can and cannot support.
Department of Statistics, University of Nebraska–Lincoln
Invited back a second time to advise graduate students on securing and making the most of internships, and on moving between academic and industry research.
Department of Statistics, University of Nebraska–Lincoln
Served on the search committee for a departmental Business Associate position — reviewing applications and interviewing candidates alongside faculty and staff.
STATCOM, University of Minnesota–Twin Cities
Helped run the student-led organisation providing pro-bono statistical consulting to non-profit and community organisations across the Twin Cities.
During my doctoral studies I also mentored junior graduate students informally on machine-learning and deep-learning projects, and I am prepared to supervise across artificial intelligence, machine learning, uncertainty quantification and applied statistics.
For the GenAI Document Validation System, DocEval.
Gold, Best Advance in Generative AI for Business Impact, for DocEval; Bronze, Best Advance in AI for Business Impact, for the enterprise RAG assistant.
American Business Awards, August 2025, and International Business Awards, November 2025 — both for DocEval.
Third Place and Best in Show.
First Place, University of Minnesota.
University of Nebraska–Lincoln. USD 8,000 per year for the first three years of the doctoral programme, awarded to recruit exceptional scholars pursuing a terminal degree.
University of Minnesota–Twin Cities. USD 5,500; sole recipient from the statistics department that year.
My research is on prediction that carries honest uncertainty — calibration, prediction intervals, and the statistical machinery behind them — applied to genomics, health, and generative AI systems. I am looking for the next place to do that work, and I would welcome a conversation with anyone building in that space.