Portfolio LinkedIn Blog Resume

Numi v4: Autonomous AI Coin Sorting, Rebuilt From Scratch

Featured On

Version 4 of my automated coin-sorting robot is complete.

Last summer I brought Numi v3 to the Colorado Springs Coin Club Show, the ANA Summer Seminar, and Stack's Bowers Professional Numismatist Program. It worked, people loved it, and I came home with a notebook full of things that were wrong with it. The past twelve months have been spent fixing them, and in the process, throwing away almost everything v3 was built on.

Here is a full demo walkthrough of Numi v4 in action. You'll see the entire process, from dumping a handful of coins into the hopper all the way through AI analysis and automated sorting.

I'll be upfront before the technical breakdown: v4 is not perfect. Not by a long shot. The real goal of this version was to find the gaps in my own knowledge and close them before version 5. Everything below is what I learned by building it.

What I Threw Away

Numi v3 was a Raspberry Pi 5 running about 14,000 lines of Python, bolted to a chassis made from more than 500 LEGO Technic pieces, with two 16-megapixel Arducam cameras looking at the coin. Python is gone, replaced with C++ written directly to the hardware.

Numi v3. Over 500 LEGO pieces and a Raspberry Pi 5

It was a wonderful proof of concept and a genuinely terrible machine. LEGO pieces flex under pressure. Every time the frame shifted a millimeter, the cameras lost focus and the coin chute lost alignment. I spent more time re-squaring the build than improving it.

Demoing v3. The conversations I had at these events shaped everything in v4.

So v4 got a full teardown. The LEGOS are gone, replaced with a 3D-printed frame I designed and printed myself. The Raspberry Pi is gone, replaced with a microcontroller about the size of a postage stamp.

I also printed a logo plaque. Not strictly necessary. Absolutely worth it.

The Life of a Coin

The clearest way to explain v4 is to follow one coin through the machine.

1. The hopper drops a coin

You pour coins into the hopper on top. A motor underneath rotates and feeds them down into a chute one at a time.

The hopper working through a load

Knowing when to stop feeding is harder than it sounds. There is no way to count coins mechanically, so Numi uses a laser rangefinder, a small sensor that fires an invisible laser at the coin stack and measures how long the light takes to bounce back. When the stack reaches about 100mm away, the hopper stops. When it drops back past 108mm, the hopper starts again. That gap between the two numbers keeps the motor from stuttering on and off.

2. A pusher slides it under the camera

A second motor pushes the bottom coin out of the chute and onto a small lit stage. The pusher then retracts until it physically stalls against a wall, which is how it finds its home position again. There is no encoder telling it where it is. It just drives backward until it can't.

v4 used a 5-megapixel camera looking straight down at the stage, ringed by white LEDs

The stage is lit by a ring of 19 white LEDs. Numi photographs the coin at 1600 x 1200 resolution.

3. It photographs both sides

Numi has one camera, so to see both faces of a coin it has to physically turn the coin over. A flipper arm rotates the coin, and the camera takes a second photograph.

The flipper rotating a coin so the camera can see the reverse.

The flip completing

Final flip and dump

This works. It also turned out to be one of my biggest mistakes, and I'll come back to it.

4. The AI analyzes the coin

Both photographs are sent together to Google Gemini (the same AI you can chat with in a web browser) along with a set of written instructions. Gemini sends back the coin's series, year, mint mark, a list of descriptive tags, and a decision about where it should go.

The dashboard mid-run

The part that surprises people the most is that the sorting rules are not code. They are written in plain English, in a text file, and I can change them without touching the robot at all. An excerpt from the actual instructions Numi sends with every coin:

"Save": Silver, Key Dates, or Semi-Key dates.

"Inspect": ANY known US Variety Candidates. Use your numismatic training to cross-reference the Series, Year, and Mint Mark; if this combination is associated with a specific valuable variety (e.g., 1955 Lincoln Cent, 2004-D Wisconsin Quarter), tag as "Variety Candidate" and set Action to "Inspect".

"Dump": Common face-value US coins, standard clad, and severely damaged coins.

"Misc.": Foreign coins, casino chips, tokens, and non-coin objects.

"Retry": Machine anomalies (Double-Feeds where images show different coins, overlapping coins, empty frames), or if BOTH Year and Mint Mark are unreadable.

If I want Numi to start pulling war nickels tomorrow, I add a sentence and go. That is a fundamentally different way to build a coin sorter than training a machine to recognize specific coins, and it is the single thing I am most convinced about after four versions.

5. Positioning the sorting carousel

Numi has five bins arranged in a ring: Dump for common coins, Save for silver and key dates, Inspect for variety candidates worth a human's attention, Misc for foreign coins and tokens, and Retry for anything the machine fumbled.

Once the verdict comes back, the carousel spins the coin cup over the correct bin and a dumper arm drops the coin in.

The carousel rotating. Magnets mark each bin position.

The carousel has no idea where it is in absolute terms. Small magnets mark each bin, and Numi counts them as they pass a sensor. If it ever loses count, it stops everything and re-finds home rather than guessing. A carousel that thinks it's over the Save bin when it's actually over Dump is worse than a carousel that stops.

The Strange Part: The Browser Is the Brain

This is the design decision I get the most questions about.

The chip inside Numi does not run the sorting process. It only knows how to execute single commands: push a coin, take a photo, spin to a bin. It has no idea what it's doing or why.

The actual sorting loop runs in a web browser tab. You open Numi's dashboard on a laptop or phone, and that web page is what decides what happens next. It tells the robot to push a coin, waits, asks for the photos, sends them to the AI, reads the verdict, and tells the robot which bin to use. One instruction at a time.

The microcontroller that runs the hardware. It executes commands. It doesn't make decisions.

The upside is enormous. At a coin show I can run the whole machine from my phone

The downside is just as real, and it's written into my own engineering notes: if the tab dies, the machine stops mid-coin.

Learning Electrical Engineering [From Scratch]

The part of this project I underestimated most was electricity.

Five motors, a camera, and a ring of LEDs all pulling power at the same time will brown out a microcontroller and reboot it mid-coin. The fix is a 24V power supply, a converter stepping it down for the electronics, and a 4700 microfarad capacitor acting as a buffer, a small reservoir that absorbs the surge when several motors start at once.

The wiring behind the panel. The 24V supply and the capacitor that keeps the chip from browning out.

I did not know how to do any of this a year ago. I know how to do it now.

The Numbers

‍

That works out to roughly 250 coins an hour. Put another way, having the AI identify a thousand coins costs about a quarter. Those figures aren't estimates. They come from the machine's own logs, which record the model used, the tokens consumed, the cost, and the response time for every single coin it has ever sorted.

The speed didn't come easily. Numi v4 started at about 60 seconds per coin. Getting to 15 seconds took weeks, and most of the gain came from overlapping steps that used to happen one after another. The hopper now reloads itself while the AI is still thinking about the previous coin.

What I Got Wrong

A mechanical flipper will eventually jam. Any moving part that has to physically manipulate a coin is a part that will someday fail on a coin that's bent, sticky, or oversized.

Use two cameras, not one. These first two lessons are really one lesson. The flipper only exists because I dropped from v3's two cameras to v4's one. I saved a camera and paid for it with a moving part that touches every single coin. That was a bad trade.

A funnel will eventually jam too. Coins bridge across the opening and stop feeding. I know this well enough that there is a routine in Numi's code whose only job is to violently shake the hopper back and forth when the laser reports no coins arriving, then try again. Writing a function to un-jam your machine is a sign you designed the wrong machine.

AI photo analysis is remarkably accurate when the photo is clear. Gemini pulls the year, mint mark, and series reliably, and the failures are almost never the AI's fault. They're the photograph's fault.

Lighting and camera angle are much harder than I expected. This was the genuine surprise of the whole project. Coins land at slightly different distances from the LED ring, and by my own measurements the brightness of the subject swings by roughly ten times between the near and far cases. Too bright and the lettering blows out to white. Too dark and the date disappears. I now have dozens of folders of test photographs, automatically scored for sharpness and glare, from weeks of chasing this. It ended with Numi metering the light on every single coin individually.

I can learn enough electrical engineering to build a reliable robot.

Bringing Numi to shops and shows is the best part. Nothing in the code is as valuable as watching a dealer's face when the machine pulls a silver quarter out of a junk bin, then hearing what they'd actually want it to do.

Where Numi Goes From Here

Version 5 is already sketched out, and it's shaped entirely by that list:

No flipper. Two cameras, one above and one below, photographing both faces at once. Faster and one less thing to jam.
No funnel. A wheel sorter instead, where coins are carried around a rotating disc rather than dropping through a hole they can bridge across.
A complete rethink of lighting and camera position. This is the one I'm most excited about, because it's the one that limits everything else. Better photographs mean better AI results, and better AI results are the entire product.

I'd rather ship an honest version 4 and tell you what broke than pretend I got it right the first time. Four versions in, the goal hasn't changed: make AI genuinely useful and genuinely approachable for people in our hobby.

If you have ideas, or you want Numi pointed at a bulk lot of your own, I'd love to hear from you. And if you see me at a show, come say hello. The machine will be running. If you want to keep up to date about Numi's latest developments, [as well as Dansco research], come check out my Instagram page.

Justin, aka Dansco Dude

Let's get in touch!

justinphinh@gmail.com