Every engagement starts with the records as they are — damaged, multilingual, and unlike anything in the training set.
Rebuilt a historical Arabic handwriting recognition workflow after the first vendor's output proved untrusted by the cataloging team. Added per-line confidence, human-in-the-loop correction, and preserved original diacritics.
Designed a time-coded transcription pipeline for decades of recorded interviews. Speaker diarization, custom vocabulary for regional dialect, and DACS-aligned finding aid integration.
Built a unified search layer across five institutions' catalogs without forcing schema migration. Each institution kept its authority files; the layer translated at query time.
Digitized fifty years of product packaging, advertisements, and internal memos. Metadata generation tuned for marketing vocabulary while preserving archival description standards.
Migrated forty years of bilingual administrative records from proprietary format to open standards. Error-tolerant OCR with cataloger review on every confidence dip.
Layout analysis for early-modern printed books with marginalia, running heads, and irregular pagination. Paleographer-reviewed output fed a critical-edition workflow.