React Native AI coin-grading app. Real-time on-device detection, automated image capture, backend grading workflows, and a full subscription + receipt validation stack. Built with and for someone who depends on it.
Built with and for Josh's father, whose sole retirement business is coins. The original problem was simple: coin grading is expert knowledge that takes decades to develop, and bad grades cost real money. The question was whether a phone camera and a model could close that gap.
The answer required solving a harder problem than it looked: coins have lighting as a quality signal — surface reflectivity is itself a grade indicator. A cleaned coin (chemically altered) looks like a high-grade coin to the untrained eye, but destroys collectible value. Counterfeit detection compounds it further. Josh and his collaborators are not professional numismatists. That's the point — non-experts building for expert users can't fake knowing what good looks like. The output either earns trust from collectors or it doesn't.
CoinHELPu — a numismatics YouTube channel that has spent years cataloguing coins across grades and eras — became a formal partner: providing domain expertise, an image corpus, edge-case feedback, and distribution to a real collector audience.
Real-time on-device coin detection using Vision Camera frame processors and a YOLOv8 Nano TFLite model — automated obverse/reverse image capture and cropping, backend AI grading workflows with polling, stale-response protection, and timeouts, plus subscriptions and receipt validation end-to-end.
Detection layer: Vision Camera frame processors run a YOLOv8 Nano TFLite model directly on-device, identifying coins in the live camera feed in real time. When a coin is detected with sufficient confidence, the app automatically captures the obverse (heads) and reverse (tails), crops to the coin boundary, and queues both images for grading.
Grading pipeline: Captured images are sent to a backend AI grading service. The client implements polling with stale-response protection — if a grading response is delayed, timed out, or returns a cached result, the app handles gracefully rather than displaying incorrect data. Timeouts are tuned per image complexity.
Output: Each graded coin receives a numeric grade, estimated market value, and condition narrative, along with a Grad-CAM overlay highlighting the surface regions that most influenced the grade. Users can also submit feedback — this feedback loop directly informs model improvement.
Monetization: Full subscription stack with receipt validation on both iOS (StoreKit) and Android (Google Play Billing). The app enforces entitlement checks at the grading step, not just at app launch.
The same problems that make coin grading hard — reliability in adversarial conditions, trust with expert users, feedback loops that improve the model, and graceful handling of uncertain outputs — are the same problems that make enterprise AI deployment hard.
Building for collectors who know when the grade is wrong is the same constraint as building for financial analysts who know when the document extraction is wrong. You can't ship something that earns trust at the demo stage and loses it in production. The model either works well enough that experts depend on it, or it doesn't.
The subscription + receipt validation stack, the feedback capture, the stale-response protection — these are production engineering decisions that prototype tools never surface. Building consumer apps forces you to solve problems that internal enterprise tools let you paper over.
The core product tradeoff was whether to rely entirely on a frontier vision model or build deterministic preprocessing around it. I chose the latter where reliability mattered: on-device detection and cropping first, orientation heuristics where useful, then backend grading and feedback loops. That sequencing — normalize inputs deterministically before touching the model — applies directly to enterprise document processing pipelines.