Last summer, I started work on building an AI-powered coin sorting robot, Numi v2. After several months of development, I paused the project because AI technology wasn't quite ready. However, by March, AI advancements convinced me it was time to pick it back up.
The past two months have been an intense sprint to complete Numi v3, just in time for the Colorado Springs Coin Club Show and the ANA Summer Seminar. I'm thrilled to announce that after hundreds of hours of work, Numi v3 is ready for its public debut!
I'll be demoing Numi at: - Colorado Springs Coin Club Show [June 27th-28th] - ANA Summer Seminar [June 28th-July 3rd]
If you can't attend, check out this video demo!
n13DanOc5hs
Why I Created Numi I'm a passionate coin collector and technologist [former technology product manager]. I believe strongly in the potential of Artificial Intelligence to positively impact our hobby. Many people still underestimate the potential benefits of AI to the hobby.
Numi v3 will initially focus on Lincoln Wheat cents, sorting them efficiently by year and mint mark.
How Numi Works Numi v3's sorting chassis is constructed from over 500 LEGO pieces, powered internally by a Raspberry Pi 5. I wrote the software in Python with help from Cursor AI, totaling approximately 14,000 lines of code.
When starting Numi, you'll first encounter a welcome page offering two operational modes: Standard Mode and Dealer Mode.
Standard Mode handles Numi's essential functions:
- Controlling LEGO motors to guide coins smoothly through the machine - Using dual cameras to capture high-resolution photos when coins enter the viewing area - Basic user interface for displaying and saving photos securely to Google Cloud AI analysis of coin images via Google Gemini 2.5 Flash (a powerful and affordable image recognition AI from Google) - Automated sorting by decade and year - Fully automated looping process for continuous sorting
Creating Standard Mode was by far the biggest challenge. I had to meticulously engineer every mechanical detail and software component from scratch. Once established, it laid the groundwork for easily adding other useful modes.
Dealer Mode, designed in collaboration with feedback from coin dealers nationwide, enables dealers to quickly identify valuable coins within bulk lots.
In Dealer Mode, you start by specifying coins of interest through a custom watch list:
Once running, Numi alerts you when it identifies coins matching your list, capturing detailed images and tracking analysis costs (approximately $6 per 10,000 coins analyzed). Numi can analyze roughly one coin every 15 seconds, or around 5,000 coins per day.
After processing, you can generate a custom Google Sheet report listing discovered coins along with image links.
Dealer Mode is configured for real-time feedback at demonstrations, helping me refine Numi based on practical insights from actual dealers.
With the core system in place, developing additional modes is straightforward. One idea I'm particularly excited about is Collector Mode, which would help collectors complete coin albums interactively. Users would see a digital album highlighting missing coins, dynamically updated as Numi sorts and identifies them.
I'm eager to hear your ideas or additional scenarios you'd find helpful!
Consider Numi v3 a working prototype and proof of concept. My main goal at upcoming events is to gather community feedback on the broader impact of automated AI coin sorting and its future potential.
This project has undergone significant evolution over the past two years. Initially, Numi v1 was a customized ChatGPT-powered coin grader. Numi v2 automated silver coin sorting from circulation. Each iteration brought valuable lessons that shaped this latest version. The goal has already remained the same though: to make AI genuinely beneficial and approachable for everyone in our hobby.
Looking ahead, I'll focus on refining Numi based on real-world tests. Collectors from around the country have reached out with interest in helping them sort through their bulk wheat hoards. I've had dealers offer to have me come in and run through their bulk wheaties to get concrete numbers on how time is saved and extra revenue generated.
Thank you for following my journey with Numi v3. I hope to meet you at the upcoming show or seminar!
"Next step: You need to program it to rotate and crop the images."
Agree, but surprisingly, it's not the highest-priority feature you might think it would be. Public AI models are excellent at reading texts, even on rotated images. The next issue I'm addressing is refining the lighting and photo quality. I'm not happy with Numi's inconsistencies.
After a chance to rest up, here is a recap of the feedback I heard from both collectors and dealers from the Colorado Springs Coin Club Show and the ANA Summer Seminar.
At the show, I hosted a table showcasing Numi's latest prototype. At Summer Seminar, I took the Advanced Coin Dealing course with Rob Obert from Gold & Coin Exchange and Seth Chandler from Witter Coin. I even got to demo Numi to the entire class, and both Rob and Seth were impressed and encouraged me to push Numi further.
The feedback from everyone was incredible and eye-opening. I've heard everything from expanding beyond wheat cents to Mercury dimes and Buffalo nickels to needing a processing speed of 10,000+ coins per day for larger shops. People were amazed by the analysis cost of just $6 for 10,000 coins.
The top request? Physical sorting capability to automatically separate valuable finds from rejected coins. While Numi doesn't have this feature yet, attendees could see past the current limitations and envision the future Numi promises.
The Advanced Coin Dealing course provided me with the most crucial insight for dealers; it's all about velocity and turnover. I specifically sought out this course because if Numi is going to succeed with dealers, I need to understand how they think and what they value. It's not enough that Numi unlocks revenue. It has to unlock it fast enough AND profitably enough to justify its time investment.
The support went beyond feedback. Multiple coin dealers offered to let me test Numi in their shops once I refine the prototype, giving me exactly the real-world data I need to prove Numi's value proposition. After months of complex engineering, hearing the consistent message of "Yeah, I think you have something here. Keep going" meant everything.
The next step is tightening Numi up to the point that I can arrange to visit a few dealers and do a trial run of Numi to gather key some key data. If the data shows that Numi unlocks enough value for dealers that it's worth their time, then I'll look at creating a 3D printed prototype.
If your target is bulk sorting for dealers, yes, image correction is a lower priority.
You probably want to automate the sort into 4 "buckets"
* Bulk, probably by denomination, but there are designs you can lift for that, * 2x2, i.e. worth putting in a 2x2 for album sales * Grading - fairly small # will fall into this (date, mm, and approx. grade) * Further investigation
That last one is the interesting category, because you only need your automation to hit 99 or 98% if you have a human backup who can effectively take the 1 or 2% of the load.
Not just things you couldn't classify, but maybe minor clips. dates with high-cherry-pick value. Etc.
Also, don't forget that, at least for now, a human is manually inserting coins into the device and can perform a pre-screen for "hum, weird, should look at that more closely".
-----Burton
50+ year / Life / Emeritus ANA member (joined 12/1/1973)
Life member: Numismatics International, CONECA
Member: TNA, FtWCC, NETCC, EveryCountry (online) coin club
Owned by three cats and a wife of 40+ years (joined 1983)
@BStrauss3 Agreed. While AI can do a lot, the final check still needs to have a human in the loop to do a final evaluation. I don't forsee AI or Numi doing everything end-to-end for quite some time.
@samoth I'm using Google Gemini Flash as my AI model and not a fine tuned/trained model. Gemini Flash's API is public so anyone can access is.
I haven't tested Gemini Flash and OpenAI o4 with a 55 DDO, but I've tested other errors and found they aren't accurate in identifying errors. Something like that would require a specially trained model. I'm focusing first on generally series, year and mint mark identification, but may revisit identifying errors later down the road.
Disclaimer: While a tremendous amount of effort goes into ensuring the accuracy of the information contained in this site, Coin Community assumes no liability for errors. Copyright 2005 - 2026 Coin Community Family- all rights reserved worldwide. Use of any images or content on this website without prior written permission of Coin Community or the original lender is strictly prohibited. Contact Us | Advertise Here | Privacy Policy / Terms of Use
Posted 06/27/2025 09:22 am
