From hobby banknote app to an enterprise AI pitch

The bottleneck: manual intake at MIF

At shows like MIF, grading companies handle a high volume of banknotes in a short window of time. Somewhere in the middle of that is the intake desk.

On paper, the process is simple. A collector walks up with a stack of notes they want graded. A staff member takes each note, identifies the country, series, catalog number, and basic attributes, then logs it into the system so it can move into the grading pipeline.

In reality, this is slow. Every extra second per note multiplies across a full day of submissions. The intake desk becomes a bottleneck. Collectors wait in line. Staff get tired, make small mistakes, and lose time correcting them. The grading company loses potential volume because the limiting factor is how fast a human can type and recognize obscure notes under time pressure.

That bottleneck is where a banknote AI starts to make sense. Not as a novelty, but as a way to move more notes through the front door without burning out the people sitting behind the desk.

Starting as a hobby app

The original idea was not an enterprise B2B product. It started as a simple question: can a phone recognize banknotes quickly enough to be useful for a collector?

The first version was a small hobby app. Point a phone at a note, get back a best guess of what is on screen. No integrations, no enterprise security, no SLAs. Just a camera, a model, and some metadata.

This is also where the constraints show up. Collectors are not studio photographers. Lighting is bad, hands shake, notes are in plastic sleeves, and nobody wants to carefully align anything if they do not have to. If an app cannot survive that, it will not survive a trade show intake desk where people are in a hurry.

The React Native + Expo camera pipeline

The front end of this system is a React Native app built with Expo, targeting both iOS and Android without separate native camera implementations.

I wrote about this earlier in Why my posts now start with five bullets.

The camera pipeline looks roughly like this:

• The user opens the app and lands directly in a camera view.
• Expo’s camera component streams frames from the device camera. The UI draws a simple overlay to suggest where the note should be, but it does not enforce it.
• When the user taps the shutter, the app captures a frame, normalizes orientation, and crops to the region that most likely contains the note. Heavy lifting like denoising or perspective correction is kept light on-device to keep the preview responsive.
• The processed image is sent to the backend for inference. The app shows a small loading state, with the goal of keeping round-trip time short enough that it feels like a live tool, not a file upload form.
• When the prediction comes back, the app shows the top match, confidence score, and a couple of alternates. For an intake desk scenario, this can be wired into a simple accept / correct flow.

On the surface, this is just a camera app. Underneath, it is a fast, opinionated data collection system that feeds an AI model and returns structured data in a way that a grading company can use.

Vision model confidence and thresholds

The core of the system is a vision model that tries to answer one question: what banknote is this?

Each inference returns a list of candidate notes with confidence scores. For example:

[
  { "id": "P-123", "confidence": 0.94 },
  { "id": "P-124", "confidence": 0.03 },
  { "id": "P-98",  "confidence": 0.02 }
]

For a hobby app, you can simply show the top result and let the user decide whether it looks right. For an intake desk with a queue of people, you need rules.

The model uses thresholds to decide what happens next:

• Above a high-confidence threshold, the app can auto-fill the intake form and let staff move on with a single confirmation tap.
• In a medium band, the app shows the top few candidates and asks the operator to pick one. This still saves time because the search space is now small.
• Below a low-confidence threshold, the app does not pretend to know. It flags the note for manual lookup so staff do not get lulled into trusting bad guesses.

These thresholds are the difference between a tool that quietly introduces errors into a grading pipeline and a tool that reduces them. Tuning them is not a one-off decision. It depends on the note universe, the lighting, the device cameras, and how much risk the grading company is willing to accept at intake.

From collector UX to intake workflow

Once an app can reasonably recognize notes in a hobby context, the next question is whether it can help at scale. That means thinking less like a collector and more like an intake operator who has to process hundreds of notes per day.

More on this in my article Productivity Hacking With AI: How I Let a Model Reorder My To‑Do List.

The workflow shifts from “one person, one phone, one collection” to “team of people, shared devices, shared queue.” That brings in boring but important questions:

• How do you associate each scanned note with a specific customer order?
• What happens when the network is unreliable on the show floor?
• How do you override the AI quickly when it is wrong, without slowing down the line?
• How do you export the intake data into the grading company’s existing systems?

None of those are solved by a nicer camera overlay. They are solved by integrating the AI pipeline into the actual business process, which is where the B2B part starts.

Talking to a major grading company

At some point, a project like this can move from side experiment to pitch. In this case, the target was the largest currency grading company in the space.

The conversation ran through direct contact with Aris at Stack's Bowers, a key player in currency and numismatics. That connection made it possible to move beyond a generic demo and discuss how a banknote AI could fit into a real intake operation at events like MIF.

The pitch is straightforward:

• There is a clear bottleneck at manual intake.
• There is a working camera and AI pipeline that recognizes banknotes on consumer hardware.
• With integration work, this can sit in front of graders and increase the number of notes accepted per day.

No promise to replace graders, no fantasy about perfect recognition. Just a proposal to automate the most repetitive part of the process and let experts focus on what actually requires expertise.

Read also Build in Public with Less Friction: My Real-Time Dev Diary Using NoteVault.

Six months from hobby to B2B pitch

The path from hobby app to pitching a market leader was about six months. The scope stayed narrow and the prototype stayed concrete.

There was a React Native and Expo app with a camera pipeline. There was a vision model producing confidence scores that could be tuned for different risk levels. There was a clear intake bottleneck at MIF that this technology could address.

Once those pieces exist, the gap between “tool for collectors” and “tool for a grading company” gets smaller. Not trivial, but small enough to justify a serious conversation with the people who move the most banknotes through their systems.