BluEclipse
☰
BluEclipse applied AI
Smart Product Capture: Scanning a Shelf Instead of Typing It
The barrier to getting a business online is rarely the website. It is the thousand products that have to be retyped into it. This is the camera workflow built to attack that directly.
- Category
- AI, computer vision and inventory automation
- Core technologies
- OCR, image processing, camera capture, AI extraction, structured data mapping
- Built for
- High-speed catalogue onboarding on Vendorah
Smart Product Capture turns a physical product into structured digital inventory using a camera, combining guided capture, OCR and AI-assisted extraction to read information straight off the packaging.
Digitising inventory is tedious in a very specific way. It is not difficult, it does not require judgement, and it never ends. For a business with hundreds or thousands of products, building an online catalogue means entering the same categories of information over and over:
- Product name
- Brand
- Price
- Description
- Weight
- Ingredients
- Allergens
- Barcode
- Variants
- Product imagery
A platform can make listing creation as pleasant as it likes; the information still has to come from somewhere. In practice that means a person sitting with the product in one hand and a keyboard under the other, transcribing a label that already contains everything they are typing.
The information is already on the box. The software should be the one reading it.
Point the camera at it instead
The core idea is exactly that simple, and most of the engineering is in making it reliable. The system guides the user through capturing the relevant sides of a product and converts what it finds into structured product data.
Products divide their information across the packaging in a fairly consistent way: the front is designed to sell, the back is designed to inform. The capture workflow treats them as two different jobs.
| Front capture | Back capture |
|---|---|
| Product name | Ingredients |
| Brand | Allergens |
| Price | Nutritional information |
| Product imagery | Manufacturer details |
| Visible packaging information | Product instructions |
| Other packaging text |
Both then combine into a single product record, which is why the workflow asks for two deliberate captures rather than one hopeful photograph.
Guided, not freehand
Capture works considerably better when the system has a say in how the image is taken. Rather than accepting arbitrary photographs, the interface guides the product into a defined capture area and can check that it is positioned properly before accepting the frame.
This is less about photography than about consistency. OCR quality is hostage to input quality, and a workflow that quietly accepts a skewed, half-lit, partially-cropped photograph has simply moved the failure further down the pipeline, where it is harder to explain.
OCR gives you words, not meaning
Optical character recognition pulls the visible packaging text out of the captured image, which removes the typing. It does not remove the thinking, because OCR has no idea what any of it is.
It will happily return 500 ml, R34.99, Vanilla and Contains milk as four strings of equal status. Which application field each belongs in is a question it was never asked and cannot answer.
That is the second layer's job:
| OCR returns | Mapped to |
|---|---|
R34.99 | Price |
500 ml | Size |
Vanilla | Variant or description |
Contains milk | Allergen information |
Once text has been extracted, AI interprets it and maps it into the structure the product system expects. The combination is what makes this useful — extraction alone gives you a block of text that still needs a human to sort, while interpretation alone has nothing to work on.
The image comes free
The same capture can also reduce the need for separate product photography: a clean portion of the front image can be extracted and prepared as the listing's primary image. One interaction then contributes both the data and the picture.
Where a barcode is available it can form part of the capture too — another way to identify a physical product, stored alongside the resulting record. The architecture treats it as a useful additional signal rather than a required field, because plenty of the products this is aimed at do not carry one.
Assisted creation, not automatic publishing
Automated extraction removes an enormous amount of typing. It does not remove the requirement that inventory information be correct, and a system that publishes whatever it thought it read is not saving anyone time — it is converting typing work into proofreading work while making the errors harder to spot.
So extracted values are presented for confirmation or correction before the product is added:
- Capture
- Extract
- Interpret
- Review
- Save
The automation does the repetitive part; the business owner keeps the final decision. It is the same division of labour the listing validator uses — the software is confident about tedium and deferential about judgement.
Value arrives with repetition
None of this is especially interesting for one product. Saving a couple of minutes on a single listing is a convenience. The case is arithmetic: two minutes saved across a thousand products is over thirty-three hours of work that does not have to happen — a hypothetical rather than a measured result, but one that scales the way catalogues actually do.
Which makes the interface's real job moving quickly from one item to the next. Capture, confirm, save, next. Anything that adds a step gets multiplied by the size of the catalogue, and so does anything that removes one.
More than OCR
Smart Product Capture is not a text-recognition feature with a camera attached. It is several systems in sequence — camera input, image processing, OCR, AI interpretation, field mapping, image extraction, human verification, inventory creation — each solving a different part of one problem.
Vendorah works with businesses carrying everything from small specialist ranges to thousands of products, and as catalogue size grows, manual entry becomes one of the largest barriers to getting started at all. This was built around that operational reality rather than as a demonstration of what OCR can do.
The goal is narrow and worth stating plainly: make adding a shelf of products feel like scanning inventory rather than doing data entry.
Built by BluEclipse. Developed for high-speed commerce onboarding.
For businesses with a physical catalogue
If the reason you are not online yet is the size of your inventory, that is the problem this was built to remove.
For technology clients
Camera workflows, OCR pipelines and language models mapped onto a real schema — computer vision wired into an application rather than demonstrated beside one.