Pratik Dahal
Statistics PhD Candidate | Data Scientist
I build statistically rigorous models and experiments, then turn them into Python/SQL systems that can be evaluated and used.
Featured Projects
Bayesian modeling, incrementality experiments, spatial statistics research, and retrieval software.
BayesWatch
Most Arkansas county-year overdose counts in CDC WONDER are suppressed. Dropping those rows would remove entire counties from the analysis.
-
Suppression-aware Negative Binomial likelihood: 362 of 525 county-years are interval-censored, and all 75 counties remain in the model.
-
Pooled and hierarchical Bayesian models in PyMC, compared on a 2018–2023 train / 2024 temporal holdout. The pooled suppression-aware model is the selected serving model; the hierarchical model is the challenger.
-
PostgreSQL pipeline with MLflow tracking, FastAPI serving, and a Streamlit dashboard with optional API-grounded explanations.
Retail Media Incrementality Platform
In a synthetic retail-media setting, attributed ROAS can credit campaigns for purchases that would have happened anyway. Budget decisions need a randomized holdout, not attribution alone.
-
Randomized ~80/20 treatment/control experiments, with conversion lift, incremental orders, and incremental revenue reported alongside attributed performance.
-
Analytic Wald intervals for conversion lift, and member-level stratified bootstrap intervals with 1,000 replicates for order and revenue effects.
-
SQL analytical marts and rule-based budget recommendations from campaign-level incremental results.
SDM-CAR
Classical CAR models fix spatial dependence in a neighborhood matrix. If that graph is wrong, both the fit and the uncertainty can be misleading.
-
Spectral filters on the graph Laplacian that recover classical CAR and allow more flexible spatial spectra.
-
Collapsed variational inference and collapsed MCMC under the same model interface.
-
Experiments on misspecification and prediction when observations are missing in blocks.
Vectra
Internal .md and .txt files are hard to search by meaning, and generated answers need to show which chunks they used.
-
Ingest markdown/text, chunk documents, and store dense embeddings in PostgreSQL with pgvector.
-
Metadata-filtered similarity search over the index.
-
FastAPI RAG that returns an answer, citations, and chunk snippets, or a fallback when retrieval is too weak.
About
I am a PhD candidate in statistics focused on applied data science and statistical machine learning. I work mainly with Python and SQL, with an emphasis on Bayesian modeling, experimentation, uncertainty quantification, and reproducible analysis.
I enjoy turning complex, ambiguous problems into clear statistical questions, evaluating solutions carefully, and building practical tools around the results.
Research
My research is in Bayesian computation for structured models: spatial dependence on graphs, variational inference, and MCMC when an approximation needs a check. SDM-CAR is the main example. It puts CAR-type covariance in the graph spectral domain, fits the same filters with collapsed VI or MCMC, and compares them under misspecification and block-missing prediction. The through-line is uncertainty that remains usable when the neighborhood graph, or the observation process, is only partly known.
SDM-CAR code
·
Technical note