Leah Rae
Genealogy

Family history is a data problem disguised as a hobby. The raw material is century-old handwritten civil registers, typewritten certificates, scanned photographs, newspaper clippings, census images, and online profiles of wildly varying reliability — heterogeneous, unstandardized, and mostly trapped on paper or behind archive viewers.

The challenge isn't storing names. It's reading documents that range from clean typescript to 19th-century cursive in a foreign bureaucratic hand; aligning each fact to the correct person when names repeat across generations and records disagree; and tracking provenance — which record supports which claim, at what confidence — so conflicting sources can be adjudicated instead of silently overwritten.

Do all of that across ~1,400 pages and 800+ people without it collapsing into an unmaintainable pile of notes — and prove the pipeline is reliable enough that its output can be trusted.

I built a fleet of purpose-built Claude agents, each defined as a structured skill file that owns exactly one job and shares a single database. One agent reads and transcribes scanned documents and photos with vision. One walks the FamilySearch pedigree and reconciles it against the live tree. One performs deep, recursive research on a single ancestor. One audits that research actually landed. One runs data-quality checks and safe fixes. New capabilities are new skills — the system grows without becoming a monolith.

The reliability of the whole thing comes from never letting a model write to the database directly. Every model-driven extraction lands in a review sidecar that a human approves; only then does a deterministic Python step mint IDs, dedupe by content hash, enforce foreign keys, store the original file, and write the records in a single transaction — with a backup taken first, every time.

The data model treats provenance as first-class: a citation links one source to one specific claim (about a person, event, or place) with a confidence level and a conflict flag. That makes "two records disagree about a birth year" a query to resolve — not a lost edit. Everything is backed by Azure Postgres/PostGIS for the structured tree and Azure Blob for original scans and timestamped backups, and surfaces as an interactive map with a page for every ancestor.

The system now holds 834 people, 199 sources, and 551 citations across 647 geocoded places, built from ~1,400 pages of documents and photos read, transcribed, and aligned to the right individuals — migrated to a normalized schema with enforced referential integrity and zero broken references.

The maternal Italian line was stuck at a genuine dead end: an ancestor with no recorded parents and zero attached sources. The research agents surfaced the family with AI handwriting OCR, then read the original 19th-century Italian government civil registers through a gated archive viewer — catching an error in the authoritative source (the recorded birth date was wrong by two years), disambiguating a same-named cousin via a marginal marriage annotation, and committing three new ancestors with eight citations, each tied to a specific register act.

This is the exact shape of a production document-AI system — multi-agent decomposition, human-in-the-loop review gates, deterministic apply-and-audit, and full provenance — applied to a domain I genuinely love. The same instincts (reading messy real-world documents, modeling trust and conflict, automating safely) transfer directly to enterprise document intelligence, data-quality, and agentic-workflow work.