Challenge
Manual categorization and digitization had become increasingly unsustainable, slow, and costly.
By implementing an Al-powered pipeline, Zomato dramatically accelerated menu processing, cut human effort by over 50%, and unlocked faster updates to their listings, enhancing both user experience and internal efficiency at scale.
Solutions
Better data meant better experiences for customers and for the restaurants trusting the platform to represent them accurately.
Built a multimodal Al pipeline using Vision-Language Models (VLMs) like BLIP and CLIP, combined with OCR and LLM post-processing, to understand both text and visual structure.
Implemented a human-in-the-loop review system for edge cases, boosting overall trust, precision, and adaptability across different regional formats.
Created multidimensional agents for layered classification and added vector search and conversational Al to improve internal access to structured menu data.
Developed a scalable data engineering framework to support continuous ingestion, processing, and normalization of multimodal content.
Results
Accurate menus, fast updates, information that reflects reality. That's the baseline users expect and the one that's hardest to deliver at scale.
Cutting human effort by over 50% wasn't the goal. It was the byproduct of finally having a system that could handle the complexity Zomato was already living with. The models and pipelines powering this will keep changing. Computer vision will improve, language models will get more capable, new formats will emerge. What stays the same is that Zomato has a partner who understands their data deeply enough to know when to upgrade, what to trust, and how to keep the system working as the technology moves underneath it.
