Less volume. More proof.

LLM extraction, RAG evidence and deterministic IMCI rules in one auditable workflow.

A multilingual clinical decision-support prototype that combines Gemma 4, retrieval over IMCI references, deterministic safety rules and a human-review boundary.

Deterministic data-quality checks first; LLM recommendations second.

A LangGraph workflow that profiles a dataset, proposes preprocessing decisions, validates transformations and keeps a human approval point before output generation.

Chronological validation, calibrated probabilities, decision cost and subgroup diagnostics.

A synthetic approval-decision laboratory that separates fit, calibration and final holdout periods, selects a cost-sensitive threshold and exposes subgroup diagnostics through reproducible reports and an API.

From clustering diagnostics to an interface a marketing team can actually explore.

An R/Shiny product that compares unsupervised methods, profiles four customer groups and makes the assumptions and cluster diagnostics inspectable.

Rolling volatility and VaR forecasts built without looking into the future.

A tested Python system for validated prices, return modelling, rolling volatility, Value-at-Risk and exception backtesting, with synthetic data and a provenance-aware download path.

Sampling algorithms made inspectable through code, variance comparisons and animation.

A Python implementation of inverse-CDF, accept-reject, stratification, importance sampling and control variates for probability and tail-quantile estimation.

One canonical feature contract from waveform generation to served cluster assignment.

A tested audio-clustering package with 170 shared features, synthetic waveforms, Kedro pipelines, API input limits, Docker and CI.

The target SNR is measured, and tuning records stay separate from evaluation records.

A repaired signal-processing experiment with correct noise scaling, observed-SNR checks, R-peak helpers and record-level evaluation boundaries.

RAII, sanitizers and multi-OS CI turn four crashing tests into a defensible implementation.

A finite-MDP C++ project repaired with vector and smart-pointer ownership, modern CMake, four passing tests, sanitizers and Linux/Windows CI.

A synthetic control benchmark with a strict non-clinical boundary.

A 20×20 tabular Q-learning environment over a simplified ODE model, evaluated on held-out seeds against a fixed synthetic schedule.

Time-aware validation and a train-only calendar baseline replace a random split.

A tested demand-forecasting pipeline with TimeSeriesSplit, a final temporal holdout, complete serialization, FastAPI and a licensed synthetic demo path.