Why I built this
Newspapers are the first draft of history, and millions of pages of America's are already scanned. The Library of Congress digitized them for its Chronicling America collection, an incredible resource. But the transcriptions come from an earlier generation of OCR, and much of the text is too garbled to search.
Modern AI models can read this old type almost perfectly. Once you can actually read the text, you can put the articles back together, find the people and places in them, and build a much better way to read and research this material: article by article, not page scan by page scan.
What it is, technically
Old News America reprocesses historical newspapers from scratch. Pages are re-transcribed with vision-language OCR, article boundaries are reconstructed from layout and text, and every article is analyzed for the people, places, organizations, and events it references (entity linking and deduplication are still in progress). Everything is indexed for full-text search and for grounded agentic research: an LLM agent that searches the archive and cites its sources.
The collections
Notes
Occasional write-ups on how the archive is built will collect here: reading 190-year-old type with modern OCR, reconstructing articles from newspaper columns, and teaching a research agent to cite its sources.
Get in touch
Interested in the archive, the OCR and reconstruction pipeline, or collaboration? Reach Jacob Jensen at jacob.elliot.jensen@gmail.com, on GitHub, or LinkedIn.