IsItGOLD?
Flutter (iOS, Android, web) · Node.js 20 on Firebase / Google Cloud · Next.js 14 + TypeScript on Vercel
Founder & Technical Lead
AI product in production that identifies precious metals in jewellery and ore from guided photos and estimates their value — with subscriptions, a support assistant and an admin dashboard.
Scan pipeline — the model for facts, code for money
- Guided photos Several shots per item; image detail chosen per shot within token budgets.
- Pre-check A low-cost model rejects unrelated photos before a credit is charged.
- Vision LLM Responses API under a strict JSON schema; physical findings only. Enum and range checks, backoff retries, fails closed instead of storing a guess.
- Hallmark RAG Grounds every read stamp in a curated reference set; the model may only cite what was retrieved.
- Value in code Deterministic value from a daily spot-price feed, rule-based confidence score, per-scan audit trail; analyser versions roll back by runtime flag.
Case in point: retrieval-augmented hallmark resolution rolling out
Problem
26.8%
of priced scans were valued as solid metal although the analysis itself pointed to plating or no gold.
What I built
A RAG step that grounds every read stamp in a curated reference set: hybrid lexical and vector retrieval (OpenAI embeddings, Firestore Vector Search), then a strict-schema LLM call that can only cite retrieved entries, with every number taken from the reference. Unambiguous stamps resolve deterministically at no model cost.
Cost and effect
≈ $0.001
added per ambiguous scan. Projected to remove gross overvaluations on an estimated 4–9% of priced scans; rolling out behind a runtime flag with a record mode to measure it.
- Grounded AI support assistant. Answers from an admin-maintained knowledge base loaded at runtime, returns structured JSON, and routes between a fast and a reasoning model. Usage governance meters tokens and enforces per-tier cost caps.
- Evaluation and testing. Live accuracy harness over a labelled image corpus (metal accuracy, value MAPE), golden fixtures, a mock vision server and 90+ backend test suites.
- Platform. Asynchronous scans on Cloud Tasks with an idempotent credit ledger, RevenueCat subscriptions, eBay listing integration (OAuth, Trading API, LLM-ranked category suggestions) and an issue-driven GitHub workflow with AI coding agents.