Fotografowałem paragon kasowy i sztuczna inteligencja automatycznie wyodrębniła każdą pozycję

The receipt was crumpled. It had been sitting in a jacket pocket for two days, folded unevenly, with a coffee ring obscuring the top corner where the store logo usually appears. A perfectly typical condition for a receipt that someone actually carried around rather than photographed the moment it was handed over. The phone camera captured it under kitchen lighting, slightly off-angle because holding a receipt flat while also holding a phone requires a third hand that does not exist. None of this mattered. After uploading the photo to scan.yeb.to, every line item appeared as structured data. Store name, date, individual products with their quantities and prices, subtotal, tax breakdown, total amount paid, and payment method. All extracted from a crumpled, coffee-stained photo taken under mediocre lighting conditions.

The experience of watching AI parse a receipt is one of those moments where the gap between expectation and result creates genuine surprise. The expectation, based on years of dealing with OCR technology, was that maybe the store name would come through, probably the total, and the individual line items would be a mangled mess of misread characters and misaligned columns. Traditional OCR works by recognizing individual characters and assembling them into text, which means it struggles with the exact conditions that describe every real-world receipt: thermal paper that fades, ink that smears, fonts that are tiny, alignment that is imperfect, and abbreviations that only make sense if you already know what was purchased. AI-powered document scanning approaches the problem differently. Instead of reading character by character, it understands the structure of a receipt as a whole. It knows that a column of numbers on the right side probably represents prices. It knows that the text at the top is likely the merchant name. It knows that a line with a date format contains a date. This structural understanding is what transforms a photo of crumpled paper into clean, usable data.

For years, the approach to receipt management was the universal one: a shoebox. Sometimes a literal shoebox, sometimes a drawer, sometimes a folder on the desk where receipts accumulated until tax season or until the pile became embarrassing enough to trigger a sorting session. The sorting session itself was always the same miserable exercise: squinting at faded text, trying to remember what "MLKCHOC 2x" meant, typing numbers into a spreadsheet one by one, and inevitably giving up halfway through because the return on effort simply was not there. The AI scanner eliminates the effort entirely. The return now exceeds the investment by a factor that makes the old manual process seem almost comically primitive.

Why Every Receipt Matters When Your Country Joins the Eurozone

Bulgaria entered the eurozone, and with that transition came a question that affects every household: what happened to prices? The official conversion rate is fixed, but the psychological and practical impact on everyday spending is far more nuanced. A loaf of bread that cost 1.50 leva now has a euro price, and the question is whether that euro price accurately reflects the conversion or whether it was rounded up, adjusted, or otherwise inflated during the transition. The only way to answer that question with any precision is to have data. Not memory, not impression, not the vague sense that things got more expensive. Actual receipts from before and after the changeover, with actual prices that can be compared item by item.

This is where the scanner transforms from a convenience tool into a genuine research instrument. Every grocery receipt, every pharmacy printout, every hardware store purchase becomes a data point in a price tracking dataset. Scan a receipt from three months before the eurozone entry and another from three months after, and the comparison is immediate: bread went from X to Y, milk went from A to B, the household cleaning products stayed roughly the same but fresh produce increased by 12%. These are the kinds of insights that are impossible to extract from a shoebox of crumpled paper but trivially easy to generate from a database of structured receipt data. The scanner at scan.yeb.to produces exactly that structured data, and the expense tracking system at receipts.yeb.to provides the platform to store, categorize, and analyze it over time.

The combination of scanning and expense tracking creates a workflow that serves both immediate practical needs and longer-term analytical purposes. In the immediate term, every scanned receipt feeds into budget categories automatically. Groceries go to groceries. Pharmacy purchases go to healthcare. Auto-rules handle the categorization based on store name, so the manual effort after the initial setup is essentially zero: scan the receipt, confirm the categorization, move on. Over weeks and months, the accumulated data reveals spending patterns that would be invisible without systematic tracking. That weekly supermarket trip that feels like it costs "about the same every week" turns out to have increased by 15% over the past six months. The pharmacy spending that seems irregular actually follows a predictable quarterly cycle. These patterns only emerge from data, and the data only exists if the receipts get scanned rather than thrown in a drawer.

The Two Second Extraction Process and What Happens Under the Surface

When a receipt photo is uploaded to the scanner, the processing pipeline handles several distinct tasks in rapid sequence. The first is image preprocessing: correcting the orientation, adjusting contrast and brightness to compensate for poor lighting, removing background noise from the surface the receipt was photographed on, and sharpening text that may have been slightly out of focus. This preprocessing step is invisible to the user but critical to the accuracy of everything that follows. A receipt photographed at a 30-degree angle on a dark wooden table looks very different to an AI model than the same receipt photographed straight-on against a white background, and the preprocessing normalizes these variations so that the extraction quality remains consistent regardless of how the photo was taken.

The extraction itself works at multiple levels simultaneously. At the document level, the AI identifies the type of receipt and its overall structure. At the section level, it identifies the header (store information), the line item area, and the footer (totals, payment method, transaction identifiers). At the line item level, it parses each row into its component fields: description, quantity, unit price, and line total. At the character level, it resolves ambiguities that traditional OCR would stumble over, like the difference between the digit "1" and the letter "l", or the digit "0" and the letter "O". Each level of processing feeds context to the next, which is why the overall accuracy is dramatically higher than character-by-character OCR: the AI knows that a value in a price column should be a number, so it resolves ambiguous characters as digits rather than letters.

The output is structured JSON data that can flow directly into any downstream application. For receipts.yeb.to, this means the expense entry is essentially pre-filled the moment the scan completes. The store name maps to a vendor. The date maps to the transaction date. The line items populate the expense details. The total becomes the amount. The user reviews, confirms, and the expense is tracked. The entire process from photographing a crumpled receipt to having a fully categorized expense entry in a tracking system takes less time than it took to type the total amount into a spreadsheet manually.

From Shoebox to Structured Data and What Changes When Every Receipt Gets Scanned

The behavioral shift that happens when receipt scanning becomes effortless is more significant than the technology itself. When scanning a receipt takes ten minutes of manual data entry, only the important receipts get scanned. Major purchases, business expenses, things that clearly need tracking. The small daily purchases, the coffee shop visits, the convenience store stops, the vending machine receipts, these all get ignored because the effort of recording them exceeds their individual significance. But personal finance is a game of aggregates, not individual transactions. The coffee that costs three euros is irrelevant on its own. The same coffee purchased 200 times a year is a 600 euro annual expense that deserves to be visible in a budget.

When scanning is fast and categorization is automatic, the threshold for "worth scanning" drops to zero. Every receipt gets scanned. The grocery store, the gas station, the pharmacy, the restaurant. The result is a complete picture of spending that reveals patterns invisible to anyone who only tracks "the big stuff." Subscription services discovered on bank statements but forgotten about. Gradual price increases at the regular grocery store that went unnoticed because each individual increase was too small to register. Seasonal spending patterns that could inform better budgeting if they were visible. All of this emerges naturally from the simple habit of scanning every receipt, a habit that only forms when the scanning process itself is frictionless.

The AI scanner at scan.yeb.to was built for exactly this kind of effortless, habitual use. No manual data entry. No correction of OCR errors. No categorization decisions that require thinking. Point the camera, upload the photo, review the extraction, confirm. A few seconds of active effort and the data is captured. Everything that follows, the budget tracking, the spending analysis, the price comparisons, the tax preparation, all of it builds on this foundation of frictionless capture. The receipt that used to be crumpled paper in a jacket pocket is now a row in a database, and that transformation is where all the value lies.