AI Coin Sorting · Three-Point Scan · Licensed Rarity Database

Every coin you toss is a lottery ticket. Now the machine reads the ticket.

A high-speed scanner that photographs each coin at three points, checks it in under a second against a licensed database of hundreds of thousands of coins, and routes the magic ones — key dates, errors, misprints, silver — into their own bins. You stop sorting by hand. The AI sorts for rarity.

▶ See the three-point scan Compare with what exists
< 1 secper coin, AI verdict
3 scanpoints per coin
300k+coins in licensed DB
5 binsautomatic routing
COIN HUNTER·OS // LIVE DEMO SCANNING 000,412
A · OBVERSE B · VARIETY C · METAL D · ROUTE
REJECT 2019-D cent · common — routed to reject bin

01 · The Scan

Three points. One verdict. Under a second.

Hand-sorting checks maybe 10 coins a minute by eye. The scanner reads each coin at three stations, fuses the readings, and routes it before the next coin even arrives.

A · OBVERSE

Design, date & mintmark

High-resolution imaging captures the design type and the date-mintmark pair. That pair is matched against the licensed database to score base rarity — year × mint × mintage.

EXAMPLE: 1931-S cent · 866K minted · rarity hit
B · REVERSE + VARIETY

Die varieties & misprints

A second pass scans the design for doubled dies, repunched mintmarks, die cracks, cuds, and design anomalies — matching against photographed variety archives like the ones in the Cherrypickers' guides.

EXAMPLE: 1955 doubled die obverse detection
C · EDGE / METAL / WEIGHT

Errors the eye almost never sees

Edge profile plus weight and metal signature catch what photography can't: off-center strikes, clipped planchets, wrong-metal strikes, and 90%-silver coins hiding among clad.

EXAMPLE: 2.5 g silver dime vs 2.27 g clad

02 · Real Example & Comparison

The first AI sorters exist. Here's what they proved — and what comes next.

In the spring of 2026, RoboCoin shipped the first consumer AI coin-sorting machines — 50 units, built to order, reviewed on camera with live tests that sorted wheat cents, caught silver, key dates, and error coins into three exit bins. It works. Now the blueprint for the real product: the same scanning idea, scaled with a licensed rarity database and a third scan point.

RoboCoin — what exists today (the live example)

SHIPPED 2026 · ~50 UNITS · POLAND-BUILT · ~$600 DEPOSIT · REVIEWED BY QUIN'S COINS
  • "Tens of neural networks" identify and sort a coin in under a second
  • Sorts pennies, nickels, dimes, quarters — designs back to the 1910s
  • Five on-camera tests: design ID, precious metals, modern rarities, old-vs-common, errors & varieties
  • Three routing bins: high value · low value · needs verification — plus reject
  • Caught a quarter variety the reviewer didn't know existed; flags S-mint business strikes automatically
  • Review noted gaps: dirty coins need the "verify" bin, some designs still misfiled — the camera rig grades quality while it sorts
// SOURCE: shop.robocoin.ai · youtube watch?v=WqnsJut-fQw

Numi 4.0 (the other live build)

ONE-MAN PROTOTYPE · v4 2026 · ~250–300/HR · GEMINI 2.5 FLASH · ~$0.25 / 1K COINS
  • Rebuilt from scratch by Justin Hinh — 3D-printed frame, C++ micro-controller, laser-sensor hopper
  • One 19-LED-ring camera with a mechanical flipper; both sides sent to Gemini 2.5 Flash with plain-English rules ("Save: silver/key dates")
  • ~250–300 coins/hr (~10–14 s/coin) — vs. the old 60 s/coin at v1
  • 5 routing bins — Dump · Save · Inspect · Misc · Retry — driven by a browser control loop
  • Cost dropped to ~$0.25 per 1,000 coins; started as a GPT-4 penny grader (2023), then silver sorter, then this
  • Known bottleneck: the single flipper arm is the weak point — needs attention about once an hour
// SOURCE: E-Sylum v29n33 · justinhinh.webflow.io · youtube 9ImLmteQgm4

The Coin Hunter stack (this blueprint)

4-POINT SCAN · LICENSED 300K+ COIN DB · ERROR ARCHIVES · 4 EXIT BINS
  • A/B/C/D four-point scan — obverse, reverse/variety, edge/metal/weight, route — instead of a single camera or flipper pass
  • Rarity verdict from a licensed commercial database (Numista / PCGS CoinFacts class) — hundreds of thousands of coins, mintages, auction records — no in-house catalogue to maintain
  • Variety matching against photographed die archives (Cherrypickers' / Variety Vista class) — sees doubled dies and repunched mintmarks the eye misses
  • Four exit bins — R · K (keeper) · S (silver) · E (error) — automatic, with a "verify" overflow when confidence is low
  • Quality reading on the same pass: strike and wear estimate, so an MS-63 key date routes higher than a worn G-4 of the same coin
  • Same economics as the first machines (sub-second per coin, hopper-fed) at fractions of hand-sorting labor — no flipper, no hourly attention
// THE UPGRADE: more scan geometry · bigger DB · no flipper · cleaner routing
🎯

Why the comparison matters

RoboCoin proved the hard part — automated AI recognition of rarity, metal, and errors at speed works today, on real coins, on camera, with a $600 deposit. The remaining gap is coverage and confidence: a machine is only as smart as its reference data. That's exactly why this blueprint licenses the giant databases instead of building one — the difference between recognizing 5 varieties a month and 5,000.

The competitive field — August 2026

Everything we found that points a camera and an AI at a coin — from a million-install phone app to one-man hardware builds. The gaps are the roadmap.

RC
RoboCoin AI coin sorting machineShips gen-1 · 2026

RoboCoin

First-gen AI sorter — tens of onboard neural nets, sub-second verdicts. Tested live by Quin's Coins across all five sorting challenges.

under 1 s/coin~$600 deposit
Nu
One-man build · v4 2026

Numi v4

Justin Hinh's autonomous sorter — laser hopper, mechanical flipper, Gemini 2.5 Flash verdicts, five bins. v1 was a 2023 GPT-4 grader.

~10–14 s/coin~$0.25 / 1,000
This blueprint

Coin Hunter

Our stack: denomination-first modular DB, dual cameras with no flipper, edge-first AI — thousands an hour at near-zero per-coin cost.

target ~3,600/hredge-first
Ry
Ryedale penny sorterEst. 2000s · pennies

Ryedale Sorter

The proven non-AI baseline — metal-signature discrimination separates copper cents at ~50+ rolls/hr. No rarity reading, no errors.

~50+ rolls/hr$500–900
CS
CoinSnap mobile app logo10M+ installs · ~2020

CoinSnap

Mobile AI identifier — snap a coin, get rarity, error candidates, and market values. Manual placement, single-photo pace.

~5–15 s/photofreemium app
Co
Coinoscope reverse image-search coin appFree · since ~2017

Coinoscope

Reverse image-search across major catalogues to identify a coin by photo. Great reference tool — no sorting, no grading.

seconds / photofree
AI
Graded coin in a PCGS slab2024 · photos only

Grading AI

eBay and PCGS-style AI condition models grade a single coin's quality (MS 60–70 scale) from a photo — effectively zero sorting.

seconds / photograde, not sort

Numi — version by version

Justin Hinh's one-person arc — from a GPT-4 coin grader that sorted pennies, to a rebuilt autonomous sorting machine, to a v5 already leaning further into sorting.

GPT
v1 · 2023

Numi v1 — ChatGPT grader

Custom GPT-4 grader inside a ChatGPT. It started with pennies — identify them, then grade them. The whole arc began here.

2023 · GPT-4 · 60 s/coin start
2v2 · 2024

Numi v2 — silver sorter

Automated silver-coin sorting from circulation, built with accessibility in mind. Paused when AI still wasn't ready for it.

2024 · silver · paused
Numi v3 LEGO chassis beside the v4 frame
v3 · Jun 2025

Numi v3 — LEGO + Raspberry Pi 5

500-piece LEGO chassis, Raspberry Pi 5, ~14,000 lines of Python (Cursor), dual cameras, Gemini 2.5 Flash, sort-by-decade wheat pennies. Demoed at the ANA.

15 s/coin · 5,000/day · $6/10k
Numi v4 rebuilt autonomous coin sorter
v4 · Aug 2026

Numi v4 — rebuilt from scratch

3D-printed frame, C++ on a micro-controller, laser-sensor hopper, 19-LED ring, one flipper + 5MP camera, plain-English Gemini verdicts, a browser carousel dropping into five bins (Dump/Save/Inspect/Misc/Retry).

~250/hr · ~$0.25 / 1,000
ProductOn market sinceSpeedAI engineCostStatus
RoboCoin (AI sorter)proto 2023 · 50 units 2026<1 s/coin · thousands per hourtens of onboard neural nets~$600 deposit, built to ordershipping gen-1, Poland
Numi v4 (Justin Hinh)v1 2023 · v4 Aug 2026~250–300/hr · ~10–14 s/coinGoogle Gemini 2.5 Flash, cloud~$0.25 per 1,000 coinsone-man prototype, 5 bins
Numi v3 (LEGO + RPi 5)June 2025~240/hr · 15 s/coin · 5,000/dayGemini 2.5 Flash, dual cameras~$6 per 10,000 coinsANA Summer Seminar demo
Numi v1–v2late 2023 · 202460 s/coin (start)GPT-4 grader → silver sorter—pennies first, then silver
CoinSnap (mobile)~2020manual · ~5–15 s/photocloud vision + value feedfreemium app10M+ installs
Coinoscope (mobile)~2017seconds per photoreverse image searchfreeactive
eBay AI grading2024seconds per photograding model (MS 60–70)free with listingphotos only — no sorting
Ryedale sorter2000s~50+ rolls/hrnone — metal signature$500–900proven, pennies only

Why ours will be faster and cheaper

Determined from watching both machines fail and stall — every lever below is a bottleneck one of them hit first.

Efficiency leverWho teaches usThe move
Denomination-first modular databaseNumi feeds the full Gemini context window every coinPre-sort by denomination, then load only that denomination's slice of the licensed catalogue. Tighter context = faster verdict, cheaper tokens, and the AI spends its attention on year/mint/variety instead of "which of 300,000 globals is this?"
Two cameras — kill the flipperNumi v4's single flipper arm weakens; after ~1 hour you have to go fix itDual fixed 5MP+ stations photograph both sides simultaneously. Halves the per-coin cycle and deletes the weakest mechanical part in the whole machine class.
Edge-first AINumi: every coin → Gemini Flash. RoboCoin: nets onboard, <1 sSmall on-device nets date/mint/denom the ~95% common coins in milliseconds at ~$0. The cloud LLM sees only the 2–5% candidates — thousands per hour at near-zero variable cost.
Light that never liesNumi v4's 19-LED ringFixed geometry, diffuse ring + cross-light, matte backdrop, per-station calibration. Clean consistent pixels beat a bigger model on dirty, worn, greasy coins every time.

03 · The Rarity Engine

What the AI is hunting for, every single coin.

Five detection channels run in parallel on every pass. Any one of them firing routes the coin out of the reject stream.

1909 Lincoln wheat cent obverse and reverseCH.01 · KEY DATES

Key dates & mintmarks

Low-mintage year–mint pairs are the bedrock of rarity. The database lookup scores each coin's mintage against its issue total and flags winners before the coin reaches the bin.

1909-S VDB (484K) $475+ · 1914-D (1.2M) $200+ · 1931-S (866K) $65+
1955 doubled die Lincoln centCH.02 · DIE VARIETIES

Doubled dies & RPMs

The 1955 doubled die was found by someone staring. Pattern-matching against photographed die archives finds duplications, repunched mintmarks (RPM), and doubled dies at machine speed.

1955 DDO $1,000+ · 1969-S DDO $25,000+ · 1972 DDO $150+
The legendary 1943 copper centCH.03 · MISPRINTS

Wrong-metal & mint errors

The scanner's third point — weight, metal signature, edge profile — exists for exactly one reason: the coins the eye can't grade by looking.

1943 copper cent $1M+ (~20 known) · 1944 steel cent · off-center strikes · clipped planchets
1943 Mercury dime — 90 percent silverCH.04 · PRECIOUS METAL

Silver & composition

Every pre-1965 dime, quarter, half, and dollar carries 90% silver. Metal-signature scanning pulls them by weight the same way the first AI sorters already do on camera.

Silver dime = 20×+ face · silver quarter = 40×+ face · every one a keeper
Close-up showing the 1955 doubled die doublingCH.05 · QUALITY

Strike & wear grade

Like the quality-grading scanner setups, the same pass estimates condition (G → MS). Two identical key dates route differently: an MS example is worth ten times the worn one.

Grade multiplies value · MS-63 key date + 10× vs G-4 · full-step / full-bell diagnostics
1888 Indian Head centCH.06 · ERA / DESIGN

Design-era & semi-keys

Not every winner is a five-figure coin. Indian Head cents, buffalo nickels, standing liberty quarters, and even certain modern issues (2009!) have real premiums — and a fixed place in the hunt.

Indian Head cent 50¢–$5 · 2009 designs up to 25× face · pre-1960 anything: check it

04 · The Database

Steal our thunder by not building what already exists.

The hard part of coin sorting was never the conveyor — it's the reference data: every design, mintmark, mintage, variety photo, and auction result. That's already been compiled, photographed, and maintained for decades, commercially. License it — a few thousand dollars a year instead of a data team for life.

Licensed coin types
300k+ Numista class
U.S. coins w/ values
39k+ PCGS CoinFacts
Reference images
700k+ CoinFacts / Variety Vista
Variety photos
10k+ die-variety archives

The scanner queries the licensed layer at scan time; the verdict updates the moment a database revises a value, with zero maintenance on your side.

  • Numista — API-licensed catalogue

    The largest free numismatic catalogue: 300,000+ collectible types worldwide with images, mintages, and composition — served through an official API.

    License & API — no maintenance
  • PCGS CoinFacts + Public API

    39,000+ U.S. coins with 700,000+ images, auction prices realized, and population data — the commercial standard for value lookup.

    Commercial license + API
  • Variety Vista / NGC VarietyPlus

    Photographed die-variety libraries — tens of thousands of doubled dies, RPMs, and repunched dates as training + matching reference for channel B.

    Reference images for AI training
  • Error-Ref / error encyclopedias

    The complete taxonomy of mint errors — off-centers, clips, wrong planchets, brockages — as structured labels for the edge/metal channel.

    Taxonomy, not guesswork

05 · The Economics

Expected value, not lottery tickets.

A scanner doesn't find you a 1909-S VDB every box — but it does convert every coin you touch into a checked, priced, routed decision. The odds below are community-reported ranges from coin roll hunting; the sliders let you use your own numbers.

Expected-value model

Tune the odds to your coins

01020
0%5%10%
01020
01530
neveraverageblessed

Odds & prices used: wheat cents $0.03 avg · pre-1960 finds $4 avg · silver coins $25 avg · key-date hits $500 avg, roughly 1 per 2,000,000 coins at normal luck — the kind of find that turns a year's sorting into a story.

Expected finds, per year
0 coins routed out
Wheat cents—
Pre-1960 / Indian Head—
Silver finds—
Key-date / error hits—
Even at average luck, a working scanner routes thousands of keepers a year — every one of them a coin you would have re-spent by hand.

06 · The Targets

The coins the scanner is dying to find.

Every entry below has been found in circulation — that's the whole point. Values are circulated-grade retail ranges from Coinflation / PCGS guides.

1909

1909-S VDB

484,000 minted
$475–550
19551955 doubled die Lincoln cent

1955 DDO

~20,000 known
$1,000+
19431943 copper cent

1943 copper

~20 known
$1,000,000+
19691969-S doubled die cent

1969-S DDO

~40 known
$25,000+
1914

1914-D

1.19M minted
$200+
1931

1931-S

866,000 minted
$65+
1950Jefferson nickel

1950-D nickel

2.6M minted
$15–40
1964Roosevelt dime

1964-D SMS dime

~300 known
$8,000+

The one-in-however-many.

Each of these has been found in circulation — that's the whole point. In 1943, roughly 20 copper-alloy cents were struck by accident instead of steel; one sold at auction for more than a million dollars. The third scan point exists so the legends have nowhere left to hide.

Values are circulated-grade retail ranges per Coinflation / PCGS guides; auction records for the legends. The licensed database layer re-prices all of this continuously — the scanner never inherits a stale value.

07 · The Silver Bonus

Even the "boring" coins can be 40× face.

While the AI hunts key dates and errors, the same stations silently pull every 90% silver coin from your loose change — dimes, quarters, halves. These are some of the real finds waiting in plain sight.

Assorted 90% silver coins90% silver stack Close-up of 90% silver coins90% close-up Pre-1964 US silver coinsPre-1964 set Silver coin close-upSilver close-up Silver dimesSilver dimes Silver coins in handIn hand Silver coins spilling from a bagSpilling out US junk silver coins montageJunk silver Stacks of coinsKeeper stacks

Every 90% silver dime carries ~2.5 g of silver, quarters ~5.6 g, halves ~11.25 g — the third scan point (edge + weight) catches them by signature at machine speed. Spot price updates from the licensed feed; the scanner never inherits a stale value.

07 · The Playbook

From bag of coins to sorted vault.

What the machine does with everything that comes out of the hopper — and what you do next.

Feed it anything

Bank boxes, estate jars, flea-market grabs, coin-store bulk bins, grandparents' dresser drawers. The scanner doesn't care about the source — only the coin. Hopper-fed, mixed denominations

Scan at A · B · C

Three stations read the obverse, the reverse and die details, and the edge/metal/weight. Every coin lands a confidence-scored verdict in under a second. <1 sec verdict · 3 scans

Route to the right bin

Key dates, errors, silver, and keepers each get their own bin. Common coins flood the reject chute back to the bank — untouched by human hands. 5 bins + verify overflow

Verify the hits

The low-confidence overflow gets a 10-second human look (microscope, scale, magnet). This is also where the AI learns — every corrected call improves the next thousand. Human-in-the-loop feedback

Grade & price

Keepers get a quality read (G→MS estimate) and a live value from the licensed database — auction prices realized included — before they ever reach a tube. Value at scan time

Sell or stack

eBay, coin dealers, collector forums, local shows. The scanner's verdict — year, mint, variety, grade, value — is the sales sheet. The data sells the coin

Key-date bin

Rarity hits from the licensed database lookup

Error bin

Varieties, misprints, off-centers, wrong metal

Silver bin

90% coins caught by weight & signature

Keeper bin

Wheat, Indian Head, pre-1960, design semi-keys

Reject chute

Common circulation coins, back to the bank

Verify overflow

Low confidence — 10 seconds of human eyes

Feedback loop

Every human correction retrains the model

Learning mode

Your corrections teach the next scan pass

08 · FAQ

Straight answers.

Does this machine already exist?
Not like this. The first AI sorters shipped in 2026: RoboCoin (robocoin.ai, ~50 units, sub-second verdicts, onboard neural nets) and Numi v4 (Justin Hinh's one-man build, ~250–300 coins/hr with Gemini 2.5 Flash, ~$0.25 per 1,000 coins, five bins — v1 was a 2023 GPT-4 grader, v2 sorted silver, v3 was a LEGO+Raspberry Pi prototype). Phone apps (CoinSnap, Coinoscope) identify single photos; eBay grades photos; Ryedale sorts pennies by metal without AI. What none of them do: the denomination-first modular database + dual-camera no-flipper + edge-first AI stack this blueprint specifies — that combination is the speed and cost gap.
Is this a real product I can buy today?
The first-gen machine: yes, build-to-order with a deposit. This blueprint's full stack: not yet — that's the point of the site. Every component is off-the-shelf: imaging conveyors, metal-sensor stations, licensed numismatic APIs (Numista, PCGS), and published error archives. The integration is the product.
How does it know a coin is rare?
It reads the date and mintmark at station A and looks them up in the licensed database — mintage figures, issue survival estimates, and population data. Rarity is scored per denomination: a 1931-S cent (866K minted) scores high; a 2019-D cent (billions) scores zero. Station B then checks for die varieties against photo archives; station C checks metal and geometry for silver and errors.
Can AI really identify errors and misprints?
Yes — that's what the first gen already demonstrated on camera (it pulled an Arizona quarter variety the reviewer had never heard of). Pattern-matching against photographed die archives is the same technology as face recognition: millions of normal coins teach it what's supposed to be there, so deviations — doubling, cracks, missing metal — stand out. The rare stuff gets a human verify before it's trusted.
Why license a database instead of building one?
Because it's already built, curated for decades, and commercially licensed: Numista (300k+ types, official API), PCGS CoinFacts (39k+ U.S. coins, 700k+ images, auction records), Variety Vista and NGC VarietyPlus (die varieties), Error-Ref (error taxonomy). Licensing costs thousands a year. Building and maintaining it is a full-time data team, forever.
What about the legal side?
Sorting, identifying, collecting, and selling coins is routine and legal. We are not describing melting or export — this is a collector's machine that routes coins to collectors' hands. (The copper-melt play is a completely separate idea — see Copper Stacker.)

09 · Sources & References

Every claim above, linked.

Quin's Coins — RoboCoin live test (37 min)Shipped AI sorter: 5 tests, 3 bins, real results + flaws RoboCoin official shop"Tens of neural networks… under a second" · built to order Numista APILargest numismatic catalogue — 300k+ types, API-licensed PCGS Public API + CoinFacts39k+ U.S. coins, 700k+ images, auction prices realized PCGS CoinFacts"The standard for the rare coin industry" — digital encyclopedia Variety VistaPhotographed die-variety archive for AI reference Error-RefComprehensive mint-error taxonomy Coinflation — Lincoln cent valuesMintages + circulated values for the key-date table PCGS price guideAuthoritative values for key dates & rarities U.S. Mint — specificationsOfficial weights/compositions (silver vs clad gap) Wikipedia — U.S. penny1943 copper legend, 1955 DDO, mintages CoinCommunity — sorting hardware threadsField reports on coin-sorting machinery E-Sylum — Numi v4: Autonomous AI Coin SortingAug 2026: laser hopper, flipper, Gemini verdicts, 5 bins, ~250/hr E-Sylum — Numi v3 prototypeJune 2025: LEGO+RPi5, 14K lines Python, dual cameras, $6/10k coins Justin Hinh — Numi project sitev1 GPT-4 grader → v2 sorter → v3/v4 autonomous builds Numi v4 — automated sorting videoDansco Dude: coin's journey, flipper, carousel, dashboard RoboCoin proto — Couch CollectiblesAug 2023 field test: silver, W-mint, die cracks, false positives CoinSnap — mobile AI identifier10M+ installs, rarity + error detection, market prices Coinoscope — mobile searchReverse image search across major catalogues

Image credits: wheat cent & doubled-die & 1943 & Mercury dime & Indian Head cents — Wikimedia Commons, PD. RoboCoin & Numi v4 stills — public product-review videos (Quin's Coins, Dansco Dude, YouTube). Ryedale photo — pennysorter.com. Graded slab — Wikimedia Commons, CC BY-SA 4.0. Monogram tiles are placeholder art for the phone-app tier. RoboCoin facts from the public review video and site. 1909 wheat cent — PD · 1955 doubled die — PD & PD · 1943 copper cent — PD · Mercury dime — PD · Indian Head cent — PD. RoboCoin facts from the public review video and roboCoin.ai. Sounds like a plan — this is a concept site; machines are illustrative.