ink-detection

Unveiling the secrets of an ancient library buried by Mount Vesuvius, this Kaggle competition, supported by the Vesuvius Challenge organization, tasked participants with detecting ink from 3D X-ray scans of charred scrolls preserved in a Roman villa in Herculaneum.

An aggregate of my kaggle competitions.

Registered human ground-truth ink evaluation for the open Vesuvius Challenge SOTA scroll data

An entire model pipeline of data ingestion, model training, evaluation, and segment management into a single CLI application for the Vesuvius Challenge.

Measure Before You Hunt — scan-quality + sheet-separability index for the 13 Grand-Prize scrolls, a validated ink instrument, and a corpus-wide screen (Vesuvius Challenge)

Trace-quality reports for Vesuvius Challenge scroll segments — flags likely segmentation errors from an ink-detection prediction (CPU-only).

SUBIT-INK replaces the standard binary ink classifier with a four-valued Belnap bilattice, enabling the model to express epistemic states beyond a simple yes/no decision.

A calibrated legibility metric over every published Vesuvius Challenge ink prediction: 46,764 equal-area windows, AUC 0.930 against blind labels, and a map of where the pipeline fails.