Statistical Data Science @ Indian Statistical Institute, Kolkata

Statistics · Machine Learning · Scientific Computing · Research Engineering

I build technical projects where the method, code, experiments, and failures can be inspected — not just summarized in a résumé bullet.

LinkedIn · Email

ACTIVE · SYNTHETIC PILOT VALIDATED · SURROGATE NOT TRAINED

A 49-case ANUGA-labelled synthetic flood-depth pilot has passed its prescribed convergence spot-checks, conservation checks, structural audit and same-environment reproducibility checks. It covers one fixed, idealized terrain; it is not field-validated flood prediction.

Current focus: production experiment design → staged data scaling → surrogate baselines. The development repository is currently private; public code and evidence links are not yet available.

The original 500-case dataset was rejected after an audit found 499 dry maps. The numerical failures and redesigned benchmark are part of the engineering record, not evidence of real-world forecasting capability.

A three-person academic team report on asymmetric fraud-classification costs. The source notebooks are not committed, so numerical results are not independently reproducible from the public repository. My documented team contribution was leading the optimization formulation and helping implement gradient descent.

EXPERIMENTAL / HISTORICAL

An experimental Hindi-language modelling project spanning data preparation, tokenization, model training, and evaluation. I retain it as evidence of the full experimental process — including where the resulting model did not perform well enough to justify inflated claims.

statistical learning     → reasoning under uncertainty
scientific ML            → models constrained by real structure
quantitative research    → hypotheses, validation, robust evaluation
LLM / model evaluation   → measuring behaviour instead of trusting demos
research engineering     → reproducible experiments and inspectable systems

For serious projects I try to make the chain visible:

problem
  ↓
assumptions
  ↓
implementation
  ↓
experiment
  ↓
result
  ↓
failure analysis
  ↓
reproduction

A project earns a prominent place here only when the evidence is strong enough to support it.

I'm moving toward stronger projects with:

• reproducible experiments,
• defensible evaluation,
• clean repositories,
• technical write-ups where they add evidence,
• and honest documentation of limitations.

Interested in quantitative research, applied ML, statistical computing, model evaluation, and research-oriented software.