BayesPilot
Track training runs and evaluations in one workflow. Model comparisons and rollout choices stay tied to the same metrics and artifacts.
Pratik Dahal
PhD candidate in statistics, building ML systems at the intersection of probabilistic inference, graph-structured models, and retrieval-backed applications.
Focus on variational inference, spatial models on graphs, and validation before production rollout.
About
I build models and the code around them: training pipelines, comparisons before deployment, and retrieval stacks where sources stay visible. Research centers on spatial and graph-structured priors when the neighborhood graph is imperfect.
Featured Projects
ML rollout tooling, RAG infrastructure, and one spatial-statistics research line.
Vector RAG Backend
Embeddings, vector search, and response wiring that keeps citations with model outputs. Tuned for production latency and operational debugging, not prototype demos.
SDM-CAR
Research. Spatial model in the CAR family with a learned spectral prior on the graph—less reliance on a fixed neighborhood matrix when the graph is imperfect. Variational and MCMC pieces for inference and prediction.
Latest Writing
Technical notes on each featured project: design rationale, tradeoffs, and follow-on work.
Designing BayesPilot: from model training to deployment decisions
Workflow for training, evaluation, and rollout in a production-oriented ML setting.
Building a production-style Vector RAG backend
Retrieval paths, latency, citation-grounded responses.
SDM-CAR explained: spectral structure in spatial Bayesian modeling
Spectral priors, misspecification, VI and MCMC for spatial prediction.
Contact
Email: mapratikdahal@gmail.com