LibriScan
Client project: Harvard Capstone for Amesbury Public Library
Technical Deep Dive
Led system design for a digitization platform that routes low-confidence words to reviewers in an automated review dashboard, with Django + Huey background processing, granular RBAC, and word-level bounding-box provenance for auditable archive exports.
Client Context
The Amesbury Public Library wanted to digitize handwritten historical manuscripts, many old, crumpled, or torn. Fully automatic OCR made too many mistakes on pages like that, and an archive cannot accept silent errors.
Execution
Built AWS Textract ingestion with confidence triage: confident words passed through and low-confidence words went to reviewers, each tied to its spot on the page image by word-level bounding boxes. Django + Huey handled background processing, with HTMX review tooling, OpenSeadragon manuscript inspection, and granular RBAC.
Outcome
Cut manual transcription effort by 90% while keeping every correction auditable through word-level provenance, delivered as the Harvard Systems Engineering master's capstone (Dean's List).
Why It Mattered
These were fragile, one-of-a-kind pages, and the library could not afford a machine quietly guessing. Automating the confident cases and handing the uncertain ones to a person showed that in high-accuracy work, a well-designed human review beats a fully autonomous model.
“Architected an AWS Textract OCR pipeline with human-in-the-loop confidence triage that cut manual transcription effort by 90% for historical manuscript digitization.”
Core Stack
Metrics
reduction
90%
workflow
Human-in-the-loop
provenance
Word-level