The inventory-counting bottleneck nobody puts on the fulfillment roadmap
Every cross-border operator I know has a spreadsheet, a Notes app entry, or a WhatsApp thread that exists purely to track “how many of X do we actually have right now?” It is the least glamorous work in the stack — below returns, below customs paperwork — and yet it is where margin quietly dies. Miscounts at the FBA prep stage become stranded inventory; miscounts at the 3PL become chargebacks; miscounts on a Temu or SHEIN restock sheet become stockouts during the exact week a listing finally ranks. So when a tool shows up claiming it can count similar objects from a photo with three taps and no model training, I pay attention — not because it is a finished enterprise product, but because it points at a category of operational drag that most sellers have simply accepted as permanent. SmartCheck is that tool, and the interesting question is not whether it works perfectly today. It is what it signals about where computer vision is heading for sellers who will never hire a data team.
What SmartCheck actually is, and the problem it picks a fight with
The pitch from maker Hiro, a master’s student at Tokyo University of Science, is deliberately narrow: upload or take a photo, tap three examples of the object you want counted, and the system finds similar objects and returns both a count and the detected positions. No app install, no sign-up required, 10 free scans to start. The setup can be saved as an “AI Template” and reused on future photos, which is the detail that matters most for anyone running recurring counts.
Why does the “three taps” framing matter? Because the incumbent approach to object counting in a warehouse or prep context is either (a) a human with a clicker, or (b) a task-specific vision model that requires collecting images, labeling data, and retraining before you can even test a new SKU. Hiro’s own framing is explicit: with task-specific AI counting, “you may need to collect images, label data, and retrain a model before you can even test a new object.” That is the real barrier. Not accuracy — friction. A seller with 400 SKUs and a rotating cast of packaging configurations cannot justify a labeling pipeline for each one.
Why this is not just “another counting app”
The honest comparison set here is not consumer counting apps. It is the low-end of industrial machine vision — the systems that live inside a Cognex or Keyence deployment — plus the DIY route of fine-tuning something like a YOLO model in a Colab notebook. Both work. Neither is accessible to a three-person DTC brand doing its own FBA prep in a garage. SmartCheck’s bet is that “good enough, instantly, in a browser” beats “accurate, eventually, with an MLOps budget” for a long tail of use cases. That is a legitimate bet, and it is the same bet that made tools like Canva beat desktop publishing suites for most small businesses.
Why Amazon sellers should care more than Shopify ones
This is the part I would push hardest on. A pure Shopify DTC brand usually has one warehouse, one 3PL, or dropship fulfillment — their counting pain is real but intermittent. An Amazon FBA seller has a compounding version of the problem: inbound shipment discrepancies against Amazon Seller Central records, FBA inventory reconciliation after Amazon’s own receiving errors, removal-order verification, and — for anyone running FBM alongside FBA — the constant question of whether the “available” number in Seller Central matches physical shelf reality. Every one of those disputes is won or lost on documentation. A photo with annotated detection positions is exactly the kind of evidence that turns “we think you lost 40 units” into a claim with a timestamp. That is a workflow Shopify-only merchants rarely need and Amazon operators need weekly.
How it differs from the tools you already pay for
Let me be specific about the competitive field, because “AI counting tool” is a crowded and vague label.
Against Helium 10 and Jungle Scout: those are demand-side tools. They tell you what to source and how a listing performs. They do not touch physical inventory. SmartCheck is supply-side and physical. No overlap, no substitution — but also no integration, which is a gap I will come back to.
Against a 3PL’s WMS: a warehouse management system like ShipBob’s or a standalone WMS tracks units by scan events. It is authoritative when the process is followed and useless when someone forgets to scan a pallet. SmartCheck is a verification layer, not a system of record. The right mental model is “spot-check tool,” not “replace the WMS.”
Against general-purpose vision models: you can absolutely upload a warehouse photo to a multimodal model and ask it to count. In my experience the failure mode is hallucinated precision — a confident number with no way to audit which objects were included. Hiro’s decision to show every detected position so you can review and correct is the single most operator-friendly design choice in the product, because it converts an untrustworthy number into an auditable one. For anyone who has ever tried to reconcile a count they did not personally perform, that distinction is the whole ballgame.
Against the “just use a spreadsheet” default: this is the real incumbent, and it is undefeated for a reason. Spreadsheets are free, flexible, and already integrated into how your ops person thinks. Any counting tool has to beat a spreadsheet on time-per-count, not on features.
What cross-border sellers can actually borrow from this
Even if you never open SmartCheck, there are three transferable lessons in how this product is positioned.
First, the “three examples” pattern is the right UX primitive for operational AI. The reason most AI tooling fails inside small e-commerce ops teams is that it demands setup before it delivers value. Klaviyo works because you can send a flow in an afternoon; the moment a tool requires a data pipeline, adoption collapses. If you are evaluating any AI vendor this quarter — for listing generation, for customer-service triage, for demand forecasting — ask one question: how many examples or how much data before I see a useful output? If the answer is “a few weeks of historical data,” you are buying a project, not a tool.
Second, showing your work beats being right. The detected-positions overlay is a small feature with an outsized trust payoff. The same principle applies to any automation you deploy: an AI-drafted refund response that shows which policy clause it cited is more usable than one that just sounds confident. Operators do not need AI to be infallible. They need it to be auditable.
Third, the GPT-6 Astra angle is worth noting for a different reason. Hiro says he used GPT-6 Astra across product and UI design, implementation, iteration, and even the launch graphics and demo video, and that without it, “turning it into a launch-ready product as a small student team wasn’t practical.” That is a one-person team shipping a functioning web app. For cross-border sellers, the implication is not “build your own tools” — it is that the cost of internal tooling has collapsed. The scrappy custom script you have been putting off because it needed a developer is now plausibly a weekend project. That changes the build-versus-buy math on the margins of your operation.
Where the math breaks
I want to be careful here, because the temptation with a tool like this is to over-extrapolate. Counting is not the bottleneck in most fulfillment operations. Receiving, putaway, picking, and packing are. If your 3PL is losing units, the loss usually happens in a process step that never had a camera pointed at it. SmartCheck helps you detect a discrepancy after the fact; it does not prevent one. That is a real but bounded value, and anyone pitching it internally as a shrinkage solution is overselling it.
Where my judgment says it falls short
The maker is unusually candid, and I will take him at his word: complex cross-sections, objects with holes, and angled photos can still cause missed or duplicate detections. For cross-border sellers, that list is not abstract. Think about the actual SKUs that are hardest to count — a bin of small accessories, a jumbled pile of apparel in poly bags, a pallet of nested items where each unit occludes the next. Those are precisely the “complex cross-sections” cases. The easy cases (a flat lay of identical boxes) are also the cases where a human count takes thirty seconds and does not need AI at all. The value curve here is inverted from where you would want it: highest accuracy where you need it least.
Then there is the integration gap. There is no mention of API access, no Amazon Seller Central integration, no WMS connector, no CSV export described. A count that lives in a browser session is a count that has to be manually transcribed into whatever system actually matters. Every manual transcription step is a place where the accuracy gain from the AI gets eaten by human error downstream. Until SmartCheck output lands directly in a spreadsheet or a WMS, its practical accuracy ceiling is set by your data-entry discipline, not its model.
Third, the pricing after the free tier is not disclosed. Ten free scans is a generous trial, but “not disclosed” on ongoing pricing is a real friction point for anyone trying to build this into a standard operating procedure. A tool you might use daily needs predictable per-seat or per-scan pricing before it can become a line item.
Fourth, and this is the structural concern: this is a student project built with heavy AI assistance, launched on Product Hunt with a demo. That is genuinely impressive, and it is also a signal about longevity. Tools in this category either get acquired, get abandoned when the maker graduates, or pivot. If you are going to build a workflow dependency on it, ask about the roadmap and the data-retention policy — especially if you are photographing anything that reveals supplier names, packaging, or volumes. A photo of your inventory is commercially sensitive. “No sign-up required” is convenient for a trial and concerning for a production workflow.
A note on the “no training” claim
Iris Carr’s comment on the launch thread nails the appeal: “The fact that there’s no object specific training needed is a pretty interesting approach. That keeps the process straightforward.” She is right, and Hiro’s response — “we wanted to remove the whole collect data → label → retrain step for each new object” — is the correct product instinct. But I would flag for operators that “no training” is a UX claim, not a technical one. The model still has to generalize from three examples, which means its accuracy is a function of how visually consistent your objects are. Three taps on three identical white boxes is a very different problem from three taps on three garments in different colors and folds. Test on your ugliest SKU, not your cleanest.
What I’d watch / test next
If you want to pressure-test this category this week without committing to anything, here is what I would do. First, pick your single most annoying recurring count — the one your ops person does every Monday — and run it through SmartCheck on the free tier, deliberately choosing the messiest photo you have. Compare the annotated output against a manual count and log the delta. Second, if it clears that bar, test the AI Template reuse on next week’s photo of the same SKU to see whether setup actually carries over. Third, and most importantly, time the full loop including transcription into your system of record — because that is the number that decides whether this beats a clicker and a spreadsheet. And watch the pricing page. A tool with no disclosed pricing and no API is a prototype, not a platform. Treat it as a signal of where operational AI is going, not as infrastructure you build on yet.
