Optic
edge
start-up
Co-founded a first-to-market app grading trading cards with AI from a phone camera. Trained on 600k+ real cards, ten-second grades, run as a live business, then sold. Still a user. Tags: Zero-to-one / AI model training / Live operations / Exit
Overview
Co-founded a first-to-market app grading trading cards with AI from a phone camera. Designed the product and the trust layer, trained the networks on 600K+ real cards, ran the live business, then sold my interest. Still a user.
Client
Own venture with partner
Role
Product Designer + Design Engineer
Timeline
2026
Platform
Mobile · iOS-first
Team
2 co-founders
The problem was a padded envelope
When the pandemic turned trading cards into a serious alternative asset class, the only way to grade one was to post it to America. Weeks to months of waiting, meaningful cost, and your most valuable possession sitting in a courier network the whole time. Grading existed to create confidence, and the process of getting it was the least confident-feeling thing a collector could do. As collectors ourselves, one in the UK and one in Sweden, my co-founder and I could see the gap precisely because we lived inside it: a whole exploding hobby locked out of understanding what their own collections were worth.
The idea was a first-to-market app with AI-powered grading: scan a card with your phone and get a graded assessment, a digital slab, live market value and a provenance record, in seconds rather than months. The positioning mattered as much as the technology. We were deliberately not a challenger to conventional grading but an evolutionary step before it, helping collectors decide whether a card was worth sending away at all. And the trust design was the real product design problem: an AI grade means nothing unless you make it legible, so I designed transparency in from the start, what the grade is, how confident the model is, where the pricing data comes from, and a provenance timeline with free regrades for life so the grade stays honest as the card ages.
Zero-to-one
2020-22
Three days to a foundation, then dogfood everything
The proof of concept was working in three days, and it became the literal foundation of the product. From that moment we used it every single day: gathering card data, stress-testing environments, lighting, hand steadiness and connection speeds, and training the neural networks on real-world conditions rather than lab-perfect scans. We set pre-launch targets and beat every one. One hundred thousand training cards planned; over six hundred thousand scanned, analysed and graded. Twenty-second grading target; ten seconds achieved. Fifty-four phone camera types verified. An eighty-one per cent grading confidence score against an eighty per cent goal. A beta community of around eighty collectors, from bedroom hobbyists to social media influencers, kept the product honest, and this was 2021: training computer vision models on messy real-world data before AI was the fashionable thing to put in a deck.
Being a co-founder meant the design job did not stop at the screens. I owned QA, the model training data and the live operation: refactoring the subscription model when server costs scaled with the collector base, moving support into a proper tooling stack and mining those support queries, which turned out to be the richest source of product insight we had. The lesson that stuck: putting genuinely new concepts into real people's hands is invaluable at every stage, and a product's success is intrinsically linked to continually listening and responding. I eventually sold my interest and the product lives on under new ownership. I still use it as a collector, which is the only exit review that really matters.
In numbers
600k+
Cards graded to train the networks
10 sec
Average grading time
3 days
Idea to working PoC