Ancestors
ancestors.myA research platform for Italian civil records, built on ~95,000 pages of 19th-century registry documents transcribed with a self-hosted vision-language OCR pipeline, then structured with NER and entity linking into browsable people, families, and acts.
- Ran a ~92K-image OCR pipeline on consumer AMD GPU hardware (ROCm/vLLM), with adaptive concurrency and failure recovery.
- Entity extraction and record linking across births, marriages, and deaths, with automated data-quality passes that cut missing-parent rates by over 90%.
- Full product UI: record viewer, persona browse, link review, and search, deployed behind nginx/gunicorn with Cloudflare.