Custom Hardware: The Card Scanning Station

At some point, taking photos of individual cards and uploading them one at a time stops being a viable workflow. When you're a vendor processing hundreds or thousands of cards for inventory, pricing, or trade shows, you need throughput. That's the problem the scanning station is designed to solve.

This post covers the design process, the hardware decisions, and the challenges of building a custom machine that feeds physical cards through a camera and identification pipeline.

Note: The scanning station is in early development and design phases. The concepts described here reflect our current direction, which continues to evolve as we prototype and test.

The Requirements

The station needs to do a few things well:

Feed cards reliably — trading cards come in different thicknesses (standard, foil, thick stock), different sizes (standard, jumbo, Japanese), and varying conditions. The mechanism can't jam, bend, or damage cards.
Capture clean images — the camera needs consistent lighting, minimal glare (especially on foils), and enough resolution to read small text like collector numbers and set symbols.
Process in real time — each card should be identified, matched to a product, and priced before the next card arrives at the camera. The entire pipeline — image capture, model inference, product lookup — needs to complete in a predictable window.
Be affordable to build — this is a bootstrapped startup. The station needs to be buildable with consumer-grade components, not industrial automation equipment.

Why Not Off-the-Shelf?

Card scanning solutions exist in the market, but they typically solve a different problem. Most focus on catalog lookup — scan a barcode or use basic image matching against a known database. They don't handle the level of variant identification we need (distinguishing a reverse holofoil from a regular holofoil from a non-holo, across every set and print run), and they don't integrate with our pricing pipeline.

We needed something that ties directly into our multi-model recognition system and our cross-marketplace pricing data. That means custom hardware.

The 3D Printing Decision

Early prototypes used off-the-shelf enclosures and brackets, and every single one required compromises — wrong dimensions for the camera position, no good mounting point for the card guide, awkward cable routing. The moment we switched to 3D-printed custom parts, iteration speed increased dramatically.

The 3D printing advantage: Design a part in CAD, print it overnight, test it the next morning, modify and reprint. A design-to-test cycle that would take weeks with machined parts takes hours. For a prototype that's still evolving, this flexibility is worth more than the dimensional precision of machined metal.

The station's structural components are all printed in-house. Material selection was an iterative process — different parts of the station have different requirements for rigidity, heat tolerance, and surface texture, especially any surface that makes direct contact with cards.

Card Feeding Mechanics

This is the hardest part of the physical design. Cards need to separate from a stack one at a time and move past the camera at a controlled speed. The challenges:

Card separation — cards in a stack stick together, especially sleeved cards or cards with textured surfaces. The feed mechanism needs to reliably isolate one card from the stack without double-feeding.
Consistent positioning — the card needs to pass the camera in a predictable position and orientation. If cards arrive at random angles, the vision models waste capacity on rotation correction instead of identification.
Non-destructive handling — these cards have real monetary value. A mechanism that scuffs edges or bends corners is worse than useless. Every contact surface needs to be card-safe.
Speed vs. reliability trade-off — faster feeding means more throughput but more jams and misfeeds. We're optimizing for reliability first, speed second.

The actuation system uses precise motor control to advance cards at consistent speed and positioning. Getting the force balance right was critical — too much and you risk damaging cards, too little and the mechanism stalls on textured or sleeved cards.

The Imaging Setup

Getting a good card image under controlled conditions is surprisingly nuanced:

The core challenge is balancing competing requirements. Foil cards need different lighting conditions than non-foils. Text on cards is small enough that resolution matters. And any environmental noise in the frame — shadows, reflections, background clutter — degrades the vision models' accuracy. We spent significant time developing a controlled imaging environment that handles these trade-offs.

Software-Hardware Integration

The station coordinates mechanical card handling with computational identification in a tight loop. The physical side feeds and positions cards; the software side captures, identifies, and prices them. The two systems need to stay in sync — the mechanical side can't advance the next card until the software confirms it's done with the current one.

The bottleneck is always inference. The speed at which the recognition pipeline processes each image directly caps the station's throughput, regardless of how fast the mechanical system can feed. This is why pipeline optimization matters — every millisecond saved on inference translates to higher throughput at the station.

Current State and What's Next

The scanning station is an active development project. The mechanical design has gone through multiple revisions, and the software pipeline integration is functional but still being refined for edge cases — unusual card sizes, severely damaged cards, and sleeved cards that affect feeding behavior.

Upcoming work includes:

Automated sorting — routing cards to different output trays based on identification results (by value tier, by condition, by set)
Batch reporting — generating inventory reports, price sheets, and collection summaries from a scan session
Multi-game support — adapting the feed mechanism for different card sizes (Magic cards are slightly different dimensions than Pokémon cards, and Japanese cards are smaller than English)
Portable version — a trade show variant that's compact enough to transport and set up at vendor tables

Building custom hardware as a software-focused startup is unconventional, but the alternative — waiting for someone else to build the exact tool we need — wasn't an option. The collector and vendor community needs high-throughput card identification, and the only way to get it is to build it.

Want to follow the hardware development or have experience with similar projects? Reach out at [email protected].