The recognition pipeline

Three models, one reliable answer

CoinVault Pro does not rely on a single black box. Visual retrieval, a large multimodal model and an on-device fallback work together so recognition is fast, accurate and works even offline.

How a scan flows

From the moment you press the shutter to a fully structured result.

01

Coin-CLIP visual retrieval

Your photo is embedded and matched against a 340,000-image reference model. This visual prefilter narrows the field to the most likely candidates before any large model is involved — fast and cheap.

02

Gemini in JSON mode

Google Gemini confirms the identification and returns a structured JSON object — type, year, mint mark, grade estimate, errors and varieties — so the result is reliable and machine-parseable, not free-form text.

03

Offline TFLite fallback

No connection? An on-device TensorFlow Lite model handles recognition locally so you can keep scanning at a coin show, an estate sale or anywhere off the grid.

Coin-CLIP retrievalGemini JSON-modeOffline TFLite fallback

Each stage hands off to the next; if the network is unavailable, recognition falls back entirely to the on-device model.

Why we built it this way

The architecture is designed for accuracy, resilience and clean data — not for a demo.

Layered, not monolithic

Each stage does what it is best at: retrieval for speed, Gemini for reasoning, TFLite for resilience. If one stage is uncertain, the chain falls back gracefully.

Structured output

Because Gemini responds in strict JSON, every field — grade, rarity, errors, value lookup keys — flows directly into grading and valuation without fragile text parsing.

Grounded in real data

Identifications are cross-referenced with the numismatic database (US Mint, Wikidata, Numista) for mintage, rarity and composition, then valued against real sold prices.

See the pipeline in action

Recognition feeds straight into grading and valuation. Try it free, then upgrade for tighter grading and deeper auction history.