Online shopping has trained us to expect price transparency.
Before buying a laptop, pair of headphones, or household item online, we can open several tabs, compare sellers, review price history, and decide whether a deal is actually good. In a physical store, that visibility largely disappears. We see one shelf price, at one location, at one moment—and usually make a decision without knowing what nearby shoppers paid for the same product.
That gap is why I built CartLens: an AI receipt scanner for local price comparison that helps shoppers answer a deceptively simple question:
Did I overpay?
CartLens scans receipts and price tags, identifies products and prices, compares them with nearby observations, and helps shoppers find less expensive local options. The larger goal is not merely to digitize receipts. It is to turn everyday purchases into useful, privacy-conscious shopping intelligence.
The internet solved online price comparison—not physical retail
Price-tracking tools work well when product pages are public, product identifiers are consistent, and historical prices can be collected repeatedly. Physical retail is different.
Local prices may vary by:
The result is what I call market darkness: the shopper can see the price directly in front of them but cannot easily see the surrounding market.
Searches such as “how to compare grocery prices between stores using receipts” or “how to find cheaper prices at stores near me” reveal the real need. People are not asking for another generic coupon feed. They want an answer based on the products they actually buy and the stores they can realistically visit.
A receipt is more than proof of purchase
A receipt is a compact record of economic activity. It may contain:
Individually, that data helps one shopper understand one transaction. Aggregated carefully and stripped of personal information, receipts can form a local price map grounded in completed purchases.
This is the core idea behind CartLens: real receipts from real shoppers can reveal the prices people actually paid in nearby stores.
That distinction matters. A listed online price may differ from an in-store price. A promotion may have expired. A delivery marketplace may include a markup. A shelf label may not reflect the final checkout total. Receipt-derived data gives the system evidence of a completed transaction.
How AI receipt scanning works
At first glance, receipt scanning looks like a standard optical character recognition problem: take an image and extract text. In practice, reliable receipt analysis is a pipeline.
1. Capture and image preparation
The system first has to deal with real-world photography: tilted paper, shadows, crumpled receipts, faded thermal printing, glare, long receipts, and cluttered backgrounds.
Image preprocessing can improve contrast, correct perspective, identify the receipt boundary, and divide long documents into manageable regions. Better input produces better OCR, but the system must still assume imperfect images.
2. OCR and layout understanding
Extracting characters is only the beginning. The application must distinguish a product line from a subtotal, tax, loyalty message, payment reference, or return policy.
Layout carries meaning. A price positioned at the far right of a product description is different from a number inside a promotion code. Receipt formats also vary by retailer, region, language, and point-of-sale system.
3. Product normalization
Retail receipts frequently use abbreviations that make sense to a store's database but not to a shopper—or another store's catalog. The same item might appear under several shortened descriptions.
CartLens therefore has to move from raw OCR text toward a normalized product identity. That can involve:
The challenge is not simply recognizing words. It is deciding whether two messy descriptions refer to the same product, a comparable product, or different products entirely.
4. Unit-price normalization
A lower sticker price does not always mean a better deal. A 12-ounce package and a 20-ounce package cannot be compared honestly without normalizing their units.
For useful grocery price comparison by receipt, a system must reason about price per ounce, pound, count, liter, or another relevant unit. It also needs to distinguish multipacks, “buy one, get one” offers, and quantity-based discounts. 5. Geospatial and temporal comparison
Price data loses value when location and freshness are ignored. A price observed hundreds of miles away may not help someone deciding where to shop today. A six-month-old observation should not carry the same confidence as one submitted yesterday.
CartLens organizes price observations around merchant location and time. The comparison layer can then prioritize nearby, recent, and sufficiently similar observations instead of presenting every price as equally relevant.
6. A verdict a human can use
The final output should not be a wall of extracted text. It should answer practical questions:
This is where an OCR utility becomes an AI shopping assistant that analyzes receipts.
From individual scans to a crowdsourced local price database One receipt can help one person. A network of verified observations can help a community.
Each useful scan can add a timestamped price observation to what CartLens calls the Geospatial Lattice—a growing map of products, merchants, locations, and observed prices. As local coverage becomes denser, shoppers can receive more relevant comparisons.
The data loop looks like this:
This network effect is especially important for physical retail, where no single public catalog captures every store, product, promotion, and regional difference.
It also creates a difficult engineering problem: more data is not automatically better data. A useful crowdsourced pricing system needs mechanisms for duplicate detection, anomaly handling, freshness decay, location validation, and confidence thresholds. If a comparison is uncertain, the product should communicate that uncertainty instead of manufacturing precision.
Why basket comparison matters more than the cheapest single item
Most people do not shop for one isolated product. They buy a basket.
Store A may have cheaper cereal, while Store B has lower prices on milk, detergent, and produce. Driving to three stores to save a few cents can waste more time and fuel than it saves.
That is why CartLens is designed to support local store price comparison for an entire shopping list, not just one headline bargain. A useful recommendation has to consider:
The best result is not always the lowest theoretical price. It is the best practical decision for that shopper.
Privacy cannot be an afterthought
Receipts can expose more than prices. They may reveal where someone shops, when they shop, what they buy, and fragments of payment information. Any receipt-analysis platform should treat that data as sensitive.
CartLens is being built around data minimization and user control. The useful part of a crowdsourced price observation is the product, merchant node, price, and time—not a person's identity or movement history.
The platform's privacy approach emphasizes keeping user choices private, avoiding the sale of SKU-level behavior to advertisers, supporting data export, and allowing users to remove personal identifiers while preserving non-identifiable price observations.
There is a broader product principle here for developers: if your AI feature depends on personal data, privacy is part of the architecture—not a policy page added after launch.
What makes CartLens different from a basic receipt scanner app?
A conventional receipt organizer answers: “What did I spend?”
CartLens is intended to answer:
That makes it a retail price intelligence platform rather than only a receipt digitizer.
It is also designed for more than groceries. The same underlying workflow can apply to household essentials, pharmacy products, beauty items, electronics, pet supplies, and other categories where local prices vary.
What I have learned while building it
Several lessons extend beyond CartLens.
AI accuracy is a product experience, not one model score
A strong OCR model can still produce a weak product if normalization, matching, or confidence handling fails. End-to-end accuracy matters more than a single benchmark.
“Same product” is a domain decision
Exact matches, equivalent sizes, generic alternatives, and substitutes are different relationships. The interface must tell users which kind of comparison they are seeing.
Freshness belongs in the data model
A price without a timestamp is incomplete. Price observations should age, and the system should reduce confidence as they become stale.
Explainability builds trust
If an app says a shopper overpaid, it should show the comparison behind that conclusion: store, observed price, distance, date, unit basis, and confidence. The shortest output may require the deepest system
“You paid a fair price” or “this item was $4 cheaper nearby” sounds simple. Producing that sentence responsibly may require computer vision, structured extraction, entity resolution, unit conversion, geospatial search, temporal weighting, and product design.
The larger opportunity: shopping intelligence for the physical economy
Retailers have sophisticated systems for pricing, inventory, demand forecasting, and customer segmentation. Individual shoppers rarely have equivalent tools.
I believe the next generation of consumer software will move beyond storing receipts and displaying charts. It will help people interpret purchases, compare realistic alternatives, detect unusual price changes, and make better decisions before the next checkout.
That is the direction behind CartLens: turn the receipt from a record of what already happened into intelligence for what to do next.
If you are interested in how to know if you overpaid for groceries, building a crowdsourced local price comparison app, or exploring how multimodal AI can make offline commerce more transparent, visit CartLens and take a look at the project. I would also love to hear from developers working on OCR, product entity resolution, geospatial systems, privacy-preserving analytics, or retail data. What do you think is the hardest technical problem in making physical-store prices genuinely transparent?