trustworthy-ml

An in-browser ML tool that replaces a single accuracy score with an honesty report: calibration, bootstrap confidence intervals, per-subgroup fairness, and selective prediction — hand-written models, no ML library, self-verified against ground truth

Certified Reduced-Lead ECG Reconstruction: a per-feature recoverability certificate with distribution-free calibration (IEEE ICASSP target)

Reliability-aware biodiversity data integration using BBS, eBird, NASA Earthdata HLS/Sentinel-2, and Prithvi/TerraMind/Clay for trustworthy ecological inference.

NLP text-classification project for deceptive-review detection, focused on trust, evaluation, and data quality.

Adversarially robust image classifier (L-inf, eps=8/255) using Friendly Adversarial Training + MART with EMA and SGDR. Public leaderboard score 0.6430.

An open-source geometric deep learning platform designed to overcome lineage confounding in antimicrobial resistance prediction through interpretable, mechanism-aware Pan-Genome Graph Neural Networks.

Shift type, not magnitude, determines ML failure modes under deployment shift — a cross-domain audit protocol and benchmark

Split-conformal prediction sets with a guaranteed marginal coverage — and a demo that naive softmax confidence plateaus at model accuracy and can't deliver 90/95% coverage

Exploratory four-probe experiment on attribution faithfulness in multi-factor LLM reasoning. Personal research notes across eight models, not a validated benchmark.

Heart disease prediction with explainability, calibration, and confidence-aware abstention.

A numpy-only reliability audit toolkit for medical-image classifiers — shortcut & leakage audits on frozen features, honest about what each does not establish.