Ancestry reposted this
Last week at the Ai4 - Artificial Intelligence Conferences, I had the opportunity to pull back the curtain on how we’re using AI at Ancestry to turn billions of historical records into meaningful discoveries for our customers. In conversation with The New Yorker's Gideon Lewis-Kraus, I walked through how our engineering teams bridge classical machine learning with modern LLM orchestration to power experiences like AncestryAI Stories. What makes this interesting to me isn’t simply the use of LLMs. It’s what happens behind the scenes to make them useful and trustworthy at scale. Our pipeline brings together billions of historical documents, from census records to WWI draft registrations, with a custom graph database, multi-model AI infrastructure, and guardrails designed to resolve complex record relationships while keeping our AI grounded in the underlying historical evidence. I also shared some of the lessons from our journey, from manual OCR and computer vision to fine-tuning open-source models and other third-party LLMs. Along the way, we’ve learned that building production-grade AI is as much about architecture, experimentation, and economics as it is about the models themselves. Ultimately, all of that engineering complexity has one purpose: helping our customers uncover stories and connections they might not have found on their own. It was great to see the response from the AI community at Ai4 and to connect with so many people interested in what it takes to move AI from experimentation into real-world products. Most importantly, I’m proud of the teams at Ancestry who are doing the hard work of making this possible every day.