Why We Started ForkFox: The Food Intelligence Story Behind the Algorithm

The Problem No One Named

Philadelphia in 2018 was a city with a restaurant scene that had moved past its own mythology but hadn't figured out how to describe itself. The cheesesteak lived in every guidebook. Pat's and Geno's were the postcards. Meanwhile, in Fishtown and West Philly and along Baltimore Avenue, restaurants were doing work that nobody had a language for yet.

Three of us kept running into the same wall. We would arrive in a neighborhood with the best intentions and a phone full of review sites. Reviews told us everything except what we actually needed to know. A restaurant could have 4.8 stars and serve mediocre food to a room full of people who had never eaten anything better. A restaurant could be unknown and be making something that worked at a level that deserved attention. The scoring felt random. The patterns invisible. The discovery process was broken, and everyone knew it, but we called it taste instead of calling it what it was: a system that had no actual way to measure what mattered.

One of us had built algorithms for logistics. One had worked in food media. One had spent a decade in restaurant operations. We kept meeting for dinner and spending more time talking about why the meal worked or didn't work than actually eating it. That's when the real question emerged: What if we could make the pattern visible? What if there was a way to score restaurants not on sentiment but on what they actually did well? Not as a replacement for knowing a place or loving a place. As a tool for knowing where to look.

Eighteen Months of Eating (and Arguing)

The first prototype was exactly as crude as you'd expect. We built a scoring framework on a Google Sheet, divided Philadelphia into grid squares, and set out to eat everywhere. No sample strategy. No pretense of scientific rigor. Just three people going to restaurants five nights a week and arguing about what made food work.

We started in neighborhoods we already knew. Fishtown. West Philly. The Italian Market. We would go to a restaurant, order what we could, and sit in the car afterward with laptops trying to articulate what we had just experienced. The first version of the framework had forty-seven variables. Flavor, yes. But also: consistency across dishes. Ingredient integrity. Value math. Context fit. How the room felt. Whether the service understood what the restaurant was trying to do. Whether the chef's vision lined up with what was actually on the plate.

Around month three, we realized we were trying to measure everything and understanding nothing. The signal was drowning in noise. We cut the framework in half. Then we cut it again. By month six, we had eighteen core attributes. By month nine, we had identified the five that actually predicted whether a restaurant would still be worth eating at in two years.

Philadelphia's food scene became our laboratory. We ate at Han Dynasty. We ate at Vetri. We ate at Tinto. We ate at counter spots on East Passyunk with no website and reservation policies that amounted to whoever showed up first. We ate Ethiopian food on Baltimore Avenue. We went back to those places eight times, twelve times, fifteen times. We built a dataset of decisions, not opinions. By the time we finished, we had visited six hundred and seventy-three restaurants across Philadelphia. We had flagged something we couldn't quite name yet: the algorithm was noticing things that the existing review ecosystem entirely missed.

What the Data Revealed (and What It Hid)

The first surprise was that consensus was not correlation. A restaurant could be beloved by the city and mediocre on the metrics that actually predicted longevity and consistency. A restaurant could be almost entirely unknown and score in the high eighties on every attribute that mattered. The algorithm was identifying things that the public discourse had no name for.

The second surprise was tougher to articulate. The framework was working. But it was working in a way that didn't feel like measurement at all. It felt like translation. We were taking the language of food—flavor, technique, integrity, intention, context—and converting it into a system that could be studied across hundreds of restaurants, across years, across neighborhoods, without losing the actual thing we cared about. Which was: Is this place real? Does it know what it's doing? Will it still be worth going to in three years?

By late 2019, we had enough data to start asking different questions. Which neighborhoods were actually overvalued? Which were undervalued? Where was the food culture moving? What did value actually mean in different parts of the city? Where was innovation happening that the media hadn't noticed yet? The data started to tell stories. Fishtown wasn't a trend. It was a structural shift in how young chefs could afford to cook. West Philly wasn't nostalgic. It was where the city's actual culinary heritage was being maintained and evolved by the communities that created it. The Italian Market wasn't retro. It was a system of provisioning that still worked better than anything the internet had invented.

And then the question nobody wanted to ask: Was our algorithm any good? We had built a system that made sense to us. But did it predict anything? Would it work in another city? Could we trust it with restaurants we had never eaten at? The answer turned out to be yes. But only if we were willing to be honest about what it actually was: not a replacement for experience, but a way to organize experience. Not a truth, but a language. A tool for seeing patterns that human attention couldn't quite hold all at once.

The COVID Test (What Broke and What Held)

In March 2020, our dataset became a history. The restaurants that closed first were not the ones we expected. The ones that survived were not always the ones that scored highest on traditional metrics. We watched **Vetri** weather through. We watched beloved institutions decide they couldn't make the mathematics work. We watched neighborhood spots that nobody had reviewed become essential infrastructure. The algorithm had never encountered a shock like this. Neither had we.

By June 2020, we were sitting with something we hadn't planned for: evidence. Which restaurants had closed? Which had pivoted and held structure? Which had become something else? We started calling restaurant owners we had eaten with. We asked: What actually kept you alive? What did you cut? What did you discover you couldn't cut without losing what made you worth saving? The answers didn't fit neatly into the framework. But they revealed something: the restaurants that had survived were the ones that had understood, before the crisis, what they were actually for.

This was the moment we realized the algorithm wasn't predictive in the way we thought. It wasn't forecasting which restaurants would succeed. It was identifying which restaurants had already made the choices that precede success. The ones with real connections to their customers. The ones where the owner understood their own value proposition. The ones where execution wasn't a feature, it was a way of being. What COVID Took: The Restaurants Philadelphia Lost and What Remains became the article we needed to write. It was also the data point we needed to survive.

By fall 2020, when we actually incorporated ForkFox, we had a completely different product than we had imagined in 2018. We weren't building a better Yelp. We were building a tool for understanding why restaurants work. Not as a black box. As a framework that a human could read, argue with, learn from. Something transparent enough that you could see what the algorithm noticed and decide for yourself whether it mattered.

From Philadelphia to Pattern

Scaling meant being honest about what we had learned in Philadelphia and what was actually transferable. The framework itself was universal. Every restaurant in every city is trying to do the same basic things: cook with integrity, serve with intention, make the economics work. But the context was entirely different. What "value" meant in Philadelphia was not what it meant in San Francisco. What "consistency" looked like in Fishtown was not what it looked like in the Mission. The algorithm could translate because it was built to understand context, not impose it.

We moved to San Francisco in early 2021. Within three months, we realized we knew nothing about San Francisco's food culture. We had to start over. We divided the city into grid squares again. We went back to eating five nights a week. We argued about what we tasted. We built a new dataset from zero. But this time, the framework held. It translated. It noticed things about San Francisco's food culture that the city's own discourse kept missing: that the tasting menu obsession had obscured the actual brilliance of the counter culture. That value wasn't about price, it was about the ratio of thoughtfulness to cost. That the city had convinced itself that eating expensively meant eating well, and that this was a choice, not a truth.

By 2022, we had enough data from two cities to see the actual pattern. The algorithm wasn't about cities at all. It was about what makes a restaurant real. And realness looked different everywhere. But it was always recognizable once you had a framework for seeing it. It was always measurable. It was always worth knowing about.

Why Transparency Matters More Than Accuracy

The biggest decision we made as a company was to never hide the algorithm. Every score we publish comes with the attributes behind it. Flavor, eighty-four. Value, ninety-two. Context, eighty-eight. Consistency, ninety. This is not how recommendation engines usually work. Usually they hide the logic to appear smarter. We did the opposite.

The reason was simple: if the algorithm is wrong, you need to know why. If a restaurant scores high on our framework but doesn't work for you, you need to be able to see exactly where we diverged. You need to be able to say: "You're right about the flavor. You're wrong about the value. You didn't understand the context the way I do." That is not a failure of the system. That is the system working exactly as designed.

This approach has cost us. Some people think the scores are too complicated. Some think we're overthinking it. Some think a restaurant should be rated on a single metric and called done. But the people who use ForkFox understand what we're actually offering: a way to think about restaurants that's more honest than sentiment and more human than a pure algorithm. A framework that notices things. Not a replacement for experience. A tool for organizing it.

We've also learned that the algorithm notices things that don't fit neatly into the stories we like to tell about food. It notices that a neighborhood can be gentrifying and still have great food. It notices that a very expensive restaurant can have mediocre execution. It notices that an unpretentious spot on a corner can understand flavor at a level that should be unthinkable at the price. It notices what we'd rather not see sometimes. And that's exactly why it matters.

What Comes Next (and What It Means)

We're in fifteen cities now. We've eaten our way through over three thousand restaurants. The framework has held. The patterns are real. The algorithm is doing the thing we built it to do: make visible what was always there but too distributed for human attention to hold.

But here's what we've learned that matters: the algorithm is not the product. Understanding is the product. We built ForkFox because we got tired of discovering restaurants the way everyone else was discovering them—through tourism boards and influencers and review sites that had every incentive to praise and almost no incentive to be honest. We built it because food deserves better than sentiment. Because restaurants deserve better than being judged on a curve that nobody understands. Because you deserve to know not just whether a place is good, but exactly why, and whether that reason matters to you.

The founding story of ForkFox is not a story about technology. It's a story about a specific moment in food culture when consensus had become so distributed and attention so fragmented that the actual signals got lost in noise. It's a story about three people who got tired of eating badly at places everyone said were great, and stopped accepting that this was normal. It's a story about building a tool that lets you see what you're looking at instead of just accepting what you're told to see.

That story started in Philadelphia. It didn't have to. But it did. And the city that built its reputation on a sandwich became the city that helped us figure out how to build something better: a way to understand what's really there.