What it is
My wife's auntie is the family genealogist. For decades she has produced a family newsletter, and the full run, 67 issues of births, weddings, obituaries, reunion write-ups, and old photographs, existed only as a pile of PDFs. All the family history was in there. None of it was findable.
Family Archive is the tool I built to change that. It reads every issue, pulls out the people, relationships, dates, and photos, and builds a searchable offline archive: a profile page for every person with their photo and life events, an interactive family tree you can pan and zoom, a search box that answers plain questions like "when did she arrive in Australia", and a GEDCOM export that loads into standard genealogy software. Every single fact links back to the newsletter issue and page it came from.
How it works
AI does the reading, not the deciding. A fleet of Claude agents processed the newsletters one issue each, extracting people and events into structured data. The final archive holds around 1,900 people, 3,300 relationships, and 350 catalogued photos.
The hard part was never the extraction. It was identity: is the "Shirley" mentioned in a 1990s issue the same Shirley from a wedding twelve years earlier? Early versions merged too eagerly and quietly invented wrong families. The fix was making the pipeline conservative and never silent. Anything ambiguous gets flagged instead of guessed, and the archive has a built-in review screen where the family expert answers Same person or Different, with her decisions folded back into the data. She stays the authority. The tool just does the reading she'd never have time for.
Privacy is built in rather than bolted on. The archive is a single HTML file that runs entirely offline, nothing uploaded anywhere, and the shareable build automatically redacts living relatives at build time. The full version travelled to its owner on a USB stick.
What I learned
AI extraction at family-history scale fails in quiet ways, and every failure taught the pipeline a rule. Names reused across generations, maiden and married names for the same woman, a casual "Don" who may or may not be Donald. The system now treats a bare first name as never enough to merge on, audits which side of a parent-child link is the parent, and runs anomaly checks after every rebuild. Each rule exists because a real error got past me first.
Status
Built and delivered, with the family review round underway. The archive itself stays private, which is why there's no live link on this one. The pipeline underneath is another story: I'm slowly shaping it into something that could do the same job for other families' shoeboxes of documents.
The bigger lesson
This is the most meaningful thing I've built with AI, and the design that made it work is the least AI-flashy part: keep a human expert in the loop and make every claim show its source. Ninety-five percent automation with visible receipts beats one hundred percent automation you can't check, especially when the subject is your own family.