How Kept Identifies Your Items
When you scan an item, your photo travels through an encrypted connection to Kept's Cloudflare Worker proxy — never directly to a third party. The proxy forwards it to GPT-5.4, OpenAI's most capable vision model, which runs two sequential AI calls:
- Call 1 — Identification: The AI reads visible details — player names, years, brand markings, set identifiers, mint marks, serial numbers, condition indicators — and returns a structured identification with a confidence level.
- Call 2 — Valuation: A second AI call estimates value across four condition tiers (Mint, Near Mint, Good, Fair/Poor) based on its training knowledge of the collectibles market.
The entire process completes in under 15 seconds. Photos are not stored by OpenAI and are not used for AI training.
Accuracy: What We're Good At, and Where We're Not
We ran a structured accuracy study across categories. These are real numbers — not marketing claims.
| Category | Avg ID Score | Avg Value Score | Avg Overall |
|---|---|---|---|
| Edge Cases | 97/100 | 97/100 | 97/100 |
| Stamps | 96/100 | 95/100 | 95.8/100 |
| Sports Cards | ~95/100 | ~72/100 | ~82/100 |
| Vinyl Records | 91/100 | 86/100 | 89.1/100 |
| Coins | ~88/100 | ~65/100 | ~75/100 |
| Toys / Figures | 84/100 | 72/100 | 79.2/100 |
| Trading Cards | 80/100 | 62/100 | 68.6/100 |
| Comics | 88/100 | 62/100 | 73.1/100 |
| Antiques | 72/100 | 45/100 | 58.0/100 |
| Art / Prints | 58/100 | 32/100 | 43.0/100 |
Best at
- Common/base versions of well-documented items (cards, coins, stamps)
- Items with clearly visible text (player names, years, mint marks)
- Recognising when it doesn't know (confidence gating)
Where accuracy varies
- Rare parallels, variants, and short prints
- Items without visible identifying text
- Niche or one-of-a-kind collectibles
- Condition estimates (lighting and angle affect results)
Your Data and Privacy
- Photos are stored in your private encrypted cloud account (Supabase) — protected by row-level security, only accessible by you
- Photos sent to the AI are used for that single scan only — not retained by OpenAI, not used for training
- Your collection data is never sold, shared, or used for advertising
- Kept is subscription-only. No ads. Ever.
- Export your full collection as CSV or PDF at any time, free on all tiers
- Delete your account and all associated data from Settings at any time
Pro Tier: Real Market Data Coming in v1.1
Free tier uses AI estimates only, drawn from GPT-5.4's training knowledge of the collectibles market.
Pro and Pro Max subscribers will receive values cross-referenced against real recent sold listings from PriceCharting and eBay — both of which track actual transaction prices for thousands of collectible categories.
All values — whether AI-estimated or market-enriched — will be clearly labeled by source so you always know which numbers come from AI training data and which come from live market transactions.
Who Built This
Kept is a solo-built app. It started because collectors — especially older ones planning their estates — deserve a private, honest tool that tells them what they own without pressuring them to sell. No marketplace, no social feed, no ads. Just a clear answer to "what do I have and what is it worth?"
Questions, concerns, or feedback: tj@kept-collections.com