A new website, Lunches.fyi, uses artificial intelligence to rank technology companies not by their market value, but by the quality of food served in their employee cafeterias. The project was created by software engineer Riley Walz in approximately one hour using voice commands with OpenAI's Codex, a task he estimated would have previously taken over 20 hours.
The site automatically scrapes publicly available menus from major tech firms, employing AI to categorise dishes and assign scores based on their perceived quality. In its initial rankings, graphics chipmaker Nvidia emerged as an unexpected leader, praised for offerings ranging from "truffle mushroom pizza" to high-quality salads.
Bug Highlights AI's Reliance on Good Data
The lighthearted project quickly underscored a fundamental principle in machine learning: an AI system's output is only as reliable as the data it processes. Amjad Masad, CEO of the software development platform Replit, publicly questioned his company's low ranking, particularly its poor score for protein content.
This prompted Walz to investigate, leading to the discovery of a coding error. "Sorry for the protein slander!" Walz stated. "Some menus didn't have nutritional facts, so the script assumed there were 0 grams of protein in all of Replit’s dishes." The flaw was caused by missing nutritional data fields in the source menus, which the AI model incorrectly interpreted as a complete absence of protein.
Results Shift After Quick Fix
Correcting this single data-handling bug immediately altered the website's rankings, demonstrating how sensitive AI-driven analyses can be to input quality. While the project is a small-scale experiment, it serves as a pointed reminder of a widespread industry challenge. "It’s a small, fun project, but a sharp reminder: AI breaks without quality data," the project's findings conclude.
The site, built using what Walz describes as "vibe coding" – a rapid, iterative development style – showcases the increasing accessibility of powerful AI coding tools. However, it also highlights the persistent need for rigorous data validation, even in seemingly simple applications.
Lunches.fyi remains live, allowing visitors to check the latest rankings of tech company cafeterias. The incident has sparked discussion among developers about the importance of building safeguards against poor or incomplete data when leveraging automated AI systems for analysis and ranking.