Tech companies have always loved competing on employee perks, and office food has become one of the clearest symbols of that race. Instead of looking only at salary and bonuses, more and more people are also comparing what actually lands on their plate. That’s the context where the Lunches.fyi project was born, an experiment that uses artificial intelligence to rank which tech companies serve the best food at work, based on public menus and structured data.
The idea is simple and at the same time pretty clever: take menus from well-known companies, run everything through AI, and turn dish descriptions into numbers, categories, and scores. The result? A very unconventional ranking, where the spotlight is not on revenue or stock performance, but on cafeteria creativity and quality. And, to many people’s surprise, one of the standout companies on the list was Nvidia, famous for dominating the AI chip market, but now also recognized for serving dishes like truffle mushroom pizza and well-crafted salads, worthy of the company’s stock performance.
Behind the fun, there is a serious point: this kind of project shows in practice how AI is completely dependent on the data it receives. When the underlying data is incomplete, skewed, or full of gaps, the algorithm will get things wrong without mercy. And that is exactly what happened in one of the most talked-about cases in this ranking, involving Replit, a tech company focused on developers.
How Lunches.fyi was born and why it drew so much attention
Lunches.fyi was created as a side project by developer and code tinkerer Riley Walz. The charm is not just in the outcome, but also in how he built the whole thing. Instead of spending hours typing every line of code, Walz used voice commands combined with Codex, OpenAI’s AI technology designed to generate and interpret code. He said that something that could have taken more than 20 hours of manual work came together in about an hour thanks to AI support.
In practice, the system works like a bot that visits public menus from tech companies, extracts the key information, and organizes it. From there, AI models kick in to classify dishes, group types of meals, understand which items are main courses, sides, desserts, and drinks, and generate a score for each cafeteria. The project’s vibe is light and almost humorous, but the technical foundation is real: scraping, automatic classification, analysis with language models, and ranking generation.
One of the aspects that drew the most attention was exactly how different the ranking turned out from what the public might expect. Instead of a big, well-known tech giant at the top, Nvidia took the spotlight with a varied menu that included more elaborate options, sophisticated pizzas, and carefully prepared greens. That combination helped build the narrative that the company’s cafeteria quality is keeping up with its strong momentum in the AI market.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
How AI reads menus and turns food into data
The engine behind Lunches.fyi is the use of artificial intelligence to do something that, until recently, would have required lots of manual analysis: understanding menus written in natural language. AI shows up in several stages of the process:
- Data collection: the system crawls public menus from tech companies and converts that information into a processable format.
- Text interpretation: language models read the names and descriptions of dishes, identifying whether something is a main dish, side, dessert, or special option.
- Classification: AI organizes meals by category, variety, estimated nutritional value, and other criteria defined in the code.
- Score generation: based on this information, the script builds ratings and rankings, comparing companies to one another.
Even when the project does not have access to full, detailed nutrition tables, it can still surface quality signals: dish diversity, the presence of lighter options, a balance between comfort food and healthier meals, variety through the days of the week, and so on. It is almost like having a hyper-fast food critic, reading dozens of menus in a short time and drawing high-level conclusions.
The use of Codex in this context is also noteworthy because it shows a very practical application of AI code models: instead of the developer having to type every function, they can just describe in natural language what they want and let the machine suggest the application skeleton. Walz described this style of work as vibe coding, a more fluid way of programming, guided by trial and error, where AI becomes a partner for rapid prototyping.
The Replit case: when bad data drags down a cafeteria’s score
The project was supposed to be just for fun, but it ended up turning into a small public lesson on the risks of trusting AI too much without checking the data. At one point, the Lunches.fyi ranking showed Replit with a very low protein score, which caught the attention of the company’s CEO, Amjad Masad. He went on social media to question the result, since according to him, the menu was not that weak on protein.
The nudge worked. Walz dug into what was going on and found a serious bug in the data pipeline: in some menus, there was no complete nutritional information. Instead of handling these empty fields with some type of warning or a more neutral default, the script simply assumed that where there was no data, there was zero protein. In other words, several Replit dishes were tagged as if they did not contain a single gram of protein, dragging the company’s average down and producing an unfair ranking.
After identifying the issue, Walz pushed a fix and published a lighthearted public apology, acknowledging the mistake. As soon as the bug was resolved, the ranking changed. Replit moved up the list, and the protein distribution started to reflect the reality of the menu more accurately. The episode became a very clear example of something many people repeat but do not always see so clearly in day-to-day life: artificial intelligence does not work miracles when the underlying data is broken.
Why this project matters more than it seems
At first glance, Lunches.fyi looks like nothing more than a bored programmer’s toy. But if you look closely, it touches on several topics at the core of today’s AI conversation:
- Data quality: a single empty field interpreted as zero was enough to distort the ranking of an entire company.
- Transparency: publicly fixing the bug shows how important it is to explain where numbers come from, especially when you are ranking people or companies.
- Prototyping speed: with tools like Codex, a project that used to take days was created in about an hour of interaction with AI.
- Lightweight AI use: not every application has to be serious or mission-critical; sometimes a playful experiment is the best way to teach complex concepts.
There is also an interesting cultural angle: the way employees, candidates, and even CEOs react to this kind of ranking says a lot about how perks are being perceived. With so many companies competing for the same talent, the day-to-day experience at the office, including what ends up on the lunch plate, can make a real difference.
AI, employee experience, and the cafeteria as a showcase
When an office food ranking starts circulating, it is not just talking about taste. It also reflects internal priorities, management decisions, and even how much a company cares about its work environment. No wonder so many people shared the results for Nvidia and other companies that ranked well.
More creative and varied menus send a few indirect messages:
- That the company is willing to invest in day-to-day well-being.
- That there is at least some concern about balancing pleasure and health.
- That there is an effort to serve different profiles and dietary restrictions.
At the same time, initiatives like Lunches.fyi show that any public data can become material for AI-driven analysis. Today it is food; tomorrow it could be other aspects of the work environment, like meeting policies, technical documentation, response times in public channels, and so on.
For people building AI products, the project serves as a practical reminder: if your application touches reputation, rankings, or public comparisons, you need to take extra care handling incomplete data. A poorly chosen default can create a completely wrong narrative, as happened with Replit’s protein score.
Limits, risks, and lessons from this kind of ranking
Even after being fixed, Lunches.fyi is still an experiment with several natural limitations. Not all menus are frequently updated, not all companies share complete menus, and in many cases, what is written does not exactly match what is actually served. On top of that, no matter how sophisticated it is, AI cannot taste, smell, or feel texture. It works with descriptions and, when available, nutritional data.
This reinforces a point that applies to almost every AI application: the result is an approximation, not an absolute truth. The best food ranking does not replace the real experience of someone who eats there every day, but it does help shed light on trends and patterns. It reveals which companies seem to invest more in culinary diversity, which ones have more conservative menus, and where nutritional balance looks more evident.
Another takeaway is the role of the community. If Replit’s CEO had not publicly questioned the low protein score, that data bug might have gone unnoticed for much longer. This interaction between end users, technical teams, and AI is a fundamental part of the improvement cycle for any system dealing with real-world data.
A future where even lunch enters the AI era
The success of Lunches.fyi shows how artificial intelligence is already mature enough to be used both in serious problems and in experimental, fun projects. While we see AI being applied in medicine, security, and major business decisions, we are also starting to see it analyze cafeteria menus, build unusual rankings, and spark new conversations about quality of life at work.
As these tools become more accessible, it is easy to imagine much more robust future versions of this kind of project: cross-referencing menus with anonymous employee feedback, analyzing waste, suggesting improvements to suppliers, or even helping people and culture teams design benefits that are better aligned with what matters to each group.
In the end, the message from this experiment is straightforward: AI is not just a serious lab thing, and bad data is still bad data, even when it goes through advanced models. The mix of creativity, critical thinking, and technology is what turns a simple idea about office food into a case study that helps a lot of people understand, in practice, how artificial intelligence works — and where it still stumbles.
