FÜD FARE

Find the product

Interpret the offer

Convert the unit

Compare with another store

Repeat

Before building Füd Fare, I compared advertised produce specials from several Chicago-area independent grocers against pricing at several nearby Aldi locations. In that exploratory sample, the local-store promotion was cheaper roughly 40% of the time.

It wasn’t a comprehensive market study, but it challenged an assumption behind the way I shopped: the larger chain wasn’t automatically offering the lowest produce price. Independent stores could be highly competitive—their offers were simply harder to discover and compare.

Füd Fare

Turn the flyer into something searchable, comparable & local

A theoretically lower price isn’t useful if reaching it costs more in time and transportation than the shopper saves.

I designed Füd Fare around a user-defined shopping radius rather than a citywide leaderboard. Shoppers choose how far they’re willing to travel, and the product limits comparisons to stores inside that area.

I deliberately started with produce because it offered a path toward comparison without requiring access to proprietary retailer SKU systems. The shopper doesn’t need to know which internal product identifier a store uses for a Bartlett pear or English cucumber. They need to know what it is, how it’s being sold, what it costs, and whether another store nearby offers meaningful savings.

That premise sounded simple.

Individual prices don’t always communicate whether something is actually a strong offer. Füd Fare therefore compares eligible prices against other listings for the same produce category within the shopper’s selected area.

I translated that comparison into the Füd Meter, borrowing the familiar visual language of airfare price trackers. Instead of asking someone to mentally compare a collection of grocery prices, the interface gives them an immediate indication of how far a current offer sits below typical advertised pricing nearby.

Normalization doesn’t replace the store’s advertised offer. Füd Fare preserves the promotion as it was presented while adding a comparable unit price when the underlying information supports one.

A shopper can therefore understand both sides of the offer: what they actually have to buy to receive the price, and how that price compares with alternatives nearby.

Finding the cheapest tomato is useful. Finding the cheapest store for the groceries you actually need is more useful.

Shoppers can add produce directly from the price directory to a running Füd List. Instead of optimizing each item independently and sending someone across the city chasing individual specials, Füd Fare evaluates the basket as a whole and identifies the store offering the strongest value across the items it can fulfill.

I deliberately avoided designing the experience around an unlimited multi-store optimization. A mathematically perfect grocery route isn’t necessarily a useful one if saving a few additional dollars requires visiting three or four different stores.

Füd Fare prioritizes the strongest single-store basket and can surface a second stop when the additional savings make it worthwhile. The goal isn’t to find the theoretical cheapest combination at any cost. It’s to help the shopper make a better real-world decision.

× Mathematically cheapest

Store A → Store B → Store C → Store D

4 stops · maximum theoretical savings

✓ Füd Fare

Your list → Store B

1 stop · strongest overall value

+ Store C — second stop only when worth it

Füd List continues into the shopping trip itself. Items can be checked off as they’re purchased, turning the same system used for price comparison into a lightweight shopping list rather than forcing the user to move between separate tools.

Where am I?

What’s cheap nearby?

What do I need?

Which store makes the most sense?

I designed Füd Fare around two parallel representations of price. The first preserves what the store actually advertised—the quantity, unit, promotion, and conditions the shopper must satisfy. The second derives a standardized price only when enough information exists to support one.

Conditional promotions follow the same principle. If a shopper must purchase four items to receive a buy-three-get-one offer, the interface can’t present the effective unit price without also preserving the quantity required to earn it.

Unit conversion solved only part of the problem. Before two prices could be compared, Füd Fare also needed to determine whether the underlying products belonged in the same comparison at all.

Grocery flyers range from extremely specific varietal names to broad labels like “apples,” “grapes,” or “cucumbers.” I developed a produce taxonomy built around micro-categories that preserve meaningful distinctions while still allowing less-specific listings to participate where the comparison remains defensible.

Apples

An unnamed apple is always apple — a cultivar is used only when the listing text names it

Nine per-cultivar micros

Produce turned out to be far more diverse than I expected. As new flyers entered the system, unfamiliar varieties and regional items continually appeared that weren’t represented in the existing taxonomy.

Rather than forcing an unknown product into the closest available category, the pipeline can flag it as a new candidate and exclude it from comparison until its classification is resolved.

An unknown product couldn’t quietly contaminate the pricing benchmark.

Not every valid grocery listing can be converted into a meaningful comparison. A store may advertise produce by crate without providing enough information to determine its weight. A package may contain an unknown number of items. A flyer may state a price clearly while leaving the quantity ambiguous.

Instead of treating normalization as all-or-nothing, I separated shopping visibility from comparison eligibility. A listing can remain available to the shopper using the store’s original advertised price while being excluded from the Füd Meter when the system can’t derive a defensible equivalent unit.

Füd Fare doesn’t generate a local benchmark from a single competing price. The Füd Meter requires multiple comparable listings using compatible units before it describes an offer as cheaper than typical in the selected area.

That constraint means some products receive a full relative-price signal while others simply display their current advertised price. The interface intentionally communicates less when the underlying dataset supports less.

The source material was inconsistent, visually dense, and sometimes ambiguous enough that a single incorrect interpretation could send a shopper to a store expecting a price that didn’t exist.

That meant extraction couldn’t simply be automated. It had to be challenged.

Flyer

Disagree

No agreement. No assumption.

Once an item has been identified, the extraction stage is deliberately narrow. The model’s job is to transcribe the facts visible in the flyer: the advertised price, quantity, unit, package information, and promotion conditions.

It does not decide what a multi-buy costs per item. It does not determine the effective price of a package. It does not calculate comparable units. Those operations are handled separately through deterministic logic.

I made that separation intentionally. A language model that performs arithmetic correctly almost every time still isn’t reliable enough for a system that may publish thousands of consumer-facing prices.

Even correctly extracted information can produce misleading results once units, quantities, and package sizes are converted. Manual auditing exposed prices that were technically explainable by the available information but obviously wrong in the context of normal grocery shopping.

Chives appeared at more than $50 per pound. Kiwi approached $60 per pound. Other products produced similarly absurd normalized values when package assumptions or units didn’t line up with the way the item was actually sold.

Chives — $56 / lb

Mathematically valid ≠ consumer valid

Every failure became another rule the next price had to survive.

The most important change wasn’t another extraction technique. It was accepting that Füd Fare didn’t need to answer every question.

If the source doesn’t clearly establish the unit, the system shouldn’t infer one. If two independent reads can’t resolve a product, it shouldn’t be forced into the taxonomy. If a package can’t be normalized confidently, it shouldn’t influence the Füd Meter.

01 — Acquire

Find and capture the current source.

Füd

What you’re shopping for.

Fare

What it costs—and what it takes to get there.

Saving money is often marketed through scarcity and pressure: limited time, lowest price, act now. But the behavior Füd Fare was designed to encourage is more positive—finding good food nearby, discovering local stores, and feeling smarter about where your grocery budget goes.

I built the visual system around that feeling. Bright teal creates a recognizable consumer signal across the interface, while oversized produce photography keeps the food itself at the center of the experience.

To give the brand a personality beyond price comparison, I developed an illustration system built around oversized produce and tiny human characters interacting with it as though the grocery aisle had become its own world.

The produce remains recognizable and desirable; the people provide the story. Everyday situations become slightly absurd—someone watering a vegetable like a garden, struggling with an enormous yam, or appearing alongside ingredients in unexpected scenes.

Keeping the characters small allowed the product to remain visually dominant while giving Füd Fare a recurring storytelling device that could expand across social, editorial, product education, and advertising.

Produce

Hero scale

Human

Secondary · ~6" scale

Everyday situation

Recognizable

One strange twist

Füd Fare

Füd Fare’s voice follows the same principle as its visual identity. Product information stays direct and understandable, while campaign work has more freedom to play with grocery language, familiar phrases, and produce-related absurdity.

That created a system where the application could remain clear and trustworthy while social content developed a much stronger personality.

Product

Clear

Local

Trustworthy

Information-first

Brand

Playful

Memorable

Human

Expressive

Füd Fare began as an experiment in whether messy, visual grocery information could be transformed into something structured enough to support a genuinely useful consumer experience.

The answer was yes—but getting there required far more than extracting text from a flyer. The system needed a product taxonomy, deterministic price logic, disagreement handling, validation rules, geographic constraints, comparison thresholds, and explicit ways to say “I don’t know.”

Early versions of the pipeline were optimized around successfully producing structured data. Auditing changed that objective. A system that returns an answer isn’t necessarily a successful system—especially when a confident error can immediately damage the user’s trust.

I became much more deliberate about separating tasks that benefit from probabilistic interpretation from tasks that require deterministic behavior. Models could help interpret messy visual information. Code should own arithmetic and business rules. Validation should challenge both. And when the evidence isn’t strong enough, the product should communicate less rather than manufacture certainty.

Before

Can the system produce an answer?

After

Has the system earned the right to show it?

The technical challenge could easily have dictated the personality of the product. I deliberately pushed in the opposite direction.

Füd Fare became one of my favorite exercises in building a consumer brand around complicated infrastructure: taking an ugly data problem and translating it into a bright, approachable experience with its own language, characters, campaigns, and visual world.

The engineering made the information usable. The brand made someone want to use it.

If I approached Füd Fare again, I’d reverse one part of the process: validate the market before validating the architecture.

I proved I could build the difficult part. The more valuable question should have come first—whether building it created enough differentiated value to warrant everything that followed.