Family History

What It Does

I have been analyzing and organizing a lot of material about my paternal grandmother’s family including audio/video interview material, scanned multi-language letters from before and during WWII, hand-typed book manuscripts and more to piece together stories and a coherent detail-filled narrative. Now several generated prose PDF documents but probably later to support material I will write like perhaps a book.

How it works

I have a large collection of old letters sent by my great-grandfather, my grandmother and her sisters and a few other family members from Europe before and during WWII. They are handwritten in mostly Yiddish in cursive Latin and Hebrew alphabets. I experimented with a variety of ways to get meaningful OCR transcriptions and translations out of frontier models. I did the same with three 200+ page volumes typed on a Hebrew typewriter and scanned 30+ years ago. I used Whisper to transcribe 40 hours of interview material and AI to catalog it. I also exported a large gedcom file of our family tree and used a slowly hand-confirmed roster to keep the people and place name together and avoid hallucination and confusion.

Built with

• hand-authored HTML rendered to PDF through headless Chrome, driven by Python build scripts that hold the corpus as structured JSON (per-letter date, writer, cache, confidence and coverage figures) and emit every edition from one stylesheet — plus a Markdown knowledge layer and CSV per-image trackers that are the actual database, with .bak-before-<change> snapshots instead of version control.
• Sources are processed locally before they reach the text: whisper.cpp (large-v3-turbo) over ffmpeg-normalized audio for ~90,000 words of interview transcript, sips/pdftoppm for scan orientation and preprocessing, pdftotext -layout with a custom artefact-repair pass for the one edition that survived only as a PDF, GEDCOM parsing across two Ancestry exports for cross-checks, and the Anthropic API for translation passes.