About Syllogeum
An open archive for ancient coin imagery
What it is
Syllogeum collects coin images and catalogue data from open sources — museum collections, institutional databases, auction records — and makes the full corpus searchable by image. Upload a photo of a coin and get the nearest visual matches across every sighting in the archive.
The goal is a permanent, accessible record of ancient coinage: the kind of thing that disappears behind dead links, paywalled databases, and auction house servers that don't get maintained. Think of it as the Internet Archive for ancient coins.
What a sighting is
A sighting is one recorded appearance of a coin — an auction listing, a museum specimen, a catalogue entry. The same physical coin can surface multiple times over decades at different auction houses, or appear in both a museum and a published catalogue. 305,000 sightings means 305,000 documented moments, not 305,000 unique coins.
The archive currently holds ~305,000 sightings across ~139,000 catalogued coin types (OCRE, CRRO, PELLA, Corpus Nummorum references). About three quarters of sightings are linked to a type card; linkage is deterministic label matching, not guesswork.
Each sighting has a stable permanent URI at syllogeum.org/sighting/{id}. These records don't move. If the original source goes offline, the sighting record and its archived image remain.
How the search works
Two embedding models — DINOv2 (from Meta) and SigLIP2 (from Google) — turn each coin image into a 768-dimensional numeric fingerprint. Every sighting gets four vectors: obverse and reverse for each model. Search compares your uploaded photo against all of them across the full corpus using cosine similarity.
The search isn't 100% accurate — no embedding model is, and ancient coins are visually similar by design. But it surfaces the right neighborhood fast, which is the hard part. Text labels and type references help narrow from there.
At current scale (~305k sightings), search runs against four FAISS indexes totaling about 3.6 GB in RAM. Each query takes under 100ms.
Data & provenance
Sightings come from open sources — the American Numismatic Society and museum collections via Nomisma linked open data, Corpus Nummorum, Europeana, and Smithsonian/Wikidata — each carrying the attribution of its origin collection. The corpus has been formally audited by license bucket, and two institutions that restrict automated access (the British Museum and the Bibliothèque nationale de France) are proactively excluded as harvest targets — honoring the spirit of their terms, not just the letter.
Full source breakdown, license buckets, what that means for reuse, the opt-out policy, corrections/takedown, and bulk downloads all live on the data & provenance page.
Dead source preservation
Syllogeum checks every source URL for liveness and submits dead pages to the Wayback Machine for archival. If a source goes offline after a sighting is recorded, the Wayback snapshot URL is stored and linked from the sighting record. The image itself is already downloaded locally, so the visual record survives regardless.
Who built it
Syllogeum is built and maintained independently by kauzmic, out of frustration that numismatic image search is still bad in 2026. Not affiliated with any institution. Not a credentialed numismatist. A meaningful portion of the code was written with AI assistance (Claude), and that's disclosed upfront.
The name comes from sylloge (Greek: συλλογή), meaning collection or anthology — as in Sylloge Nummorum Graecorum, the long-running series of scholarly coin catalogues.
How to cite
A machine-readable CITATION.cff is included in the source repository. Individual records should be cited by their stable URI (/sighting/{id}) plus the original collection named in the record's attribution line.
Contact
Questions, corrections, collaboration, takedown requests: [email protected]
Infrastructure
Self-hosted on dedicated hardware: 64 GB RAM, NVIDIA GPU, 1 TB SSD. The archive runs as a single Uvicorn/FastAPI process behind a Cloudflare tunnel. SQLite in WAL mode handles concurrent reads without locking issues at this scale.
No cloud. No subscriptions. No analytics. The server is a physical machine in a house.