There are databases where Classicists have done this manually for specific regions, Trismegistos https://www.trismegistos.org/ and Latin Inscriptions of the Roman Empire (LIRE) https://pure.au.dk/portal/en/publications/latin-inscriptions... are two major efforts I found. But there doesn't seem to be a project that did what I set out to do, although I have read in some places that it was believed to be possible.
I am not a classicist or a web developer, but I have Claude and Gemini and I can sort of read basic Latin - so I set to work. I used LIRE and another database as ground truth and built a pipeline to extract and process the inscriptions to recover the names. The process I developed uses a high end LLM like Sonnet or Gemini Pro to supervise the extraction and tuning process on a regional basis until the obvious error rate is reasonable. For this, so far, reasonable to me means less than 1-2% in the smaller initial samples of 100-500 and no observed systemic issues. The different regions often need different prompts, so this basically became an exercise in letting the higher level AI tune the prompt for the lower level AI. The extraction when measured against LIRE produces an F1 score between 0.64 and 0.87, but take this with a grain of salt.
Once I had done a few regions, I wanted to see the work, so I threw together a pretty crude website but as I am not a web developer, it was crude in how it accessed its data. It does look cool and I also added summarization, and machine translation to each entry. I wanted to eventually get feedback from an actual team of classicists and make the website work better, so I am rewriting it as we speak but it is broadly functional now with a few extra bugs but substantially improved performance compared to the old one. All entries link back to the proper sources, and the old web app linked to several additional sources where the data was present, but I haven't gotten that working again just yet on the new one. (The old web interface is still available at https://roman-names.com, but I will warn you it is clunky and not mobile friendly at all)
Key findings so far:
AI supervised AI extraction saved me time. I was manually tuning things for a while and then the runbook became an idea that I feed my instructions in and let the big AI go with sparse oversight from me.
The extraction improved significantly (by about 10 F1 points) when I fed the model the raw text including the markers, vs a cleaned up version of the text.
I just thought it was a cool little project and wanted to share. If you happen to work in any adjacent space and there is something I could do better etc let me know.