80%
Reduction in monthly costs
50%
Faster database queries
$5.7M
Projected savings over three years
Challenge
Restaurant menus are designed for diners, not for data systems.
A single complex menu could take hours.
Technomic tracks menu offerings across thousands of restaurant brands to power competitive analysis, benchmarking, and trend forecasting. Before any of that analysis could happen, an analyst had to open each menu — a PDF, a printed scan, a screenshot from an online ordering page — and hand-enter every item, price, description, category, and section.
Stylized fonts, multi-column layouts, embedded images, and structures that change completely from one concept to the next. Traditional OCR and parsing tools couldn't handle that reliably, so the work kept falling back on people.
As volume grew, the manual approach started to buckle. A single complex menu could take hours. Different analysts made different judgment calls about categorization, so consistency was a constant struggle.
The variety made it harder still. A wine list needs region and vintage. A prix fixe menu bundles pricing in a way that doesn't fit item-by-item. Kids menus, beverage catalogs, and online ordering layouts each follow their own logic.
Approach
The obvious solution was one extraction model for every menu. We recommended against it.
Forcing a single model across formats this different produces output nobody trusts, which means analysts end up reviewing every line, and the manual work returns through the back door. What looked like the faster build would have relocated the problem rather than removed it.
We recommended a system that reads each document's structure first, then processes each section with logic built for that section type. More work up front, and the only version that removes the human review step for good.
Solutions
MenuVerse AI processes menus automatically and outputs clean, validated, analysis-ready data. It runs inside Technomic's existing Azure DevOps workflow, so it fits how the team already operates.
Automated workflow triggering
Two-phase extraction
Section-specific schemas
Hybrid vision and OCR
Schema validation
Automated ticket management
TECHNOLOGY STACK
Cloud Orchestration
Azure Functions (Flex Consumption)
Queue & Triggering
Azure Queue Storage
Workflow Integration
Azure DevOps API
Database
Azure Cosmos DB
AI Models
Google Gemini 2.5 Pro, Azure OpenAI GPT-4.1, GPT-o3
OCR Engine
Tesseract, PaddleOCR
Schema Validation
Pydantic v2 (Strict Mode)
PDF Processing
PyPDF, PyMuPDF
Runtime
Python 3.13, Azure Functions Core Tools
Results
One of the team's most time-consuming workflows became largely hands-off, freeing analysts for the work that actually needs human judgment.
Faster processing
Menus that took hours of manual work now complete automatically in minutes.
Consistent outputs
Schema-driven extraction replaced individual analyst interpretation, removing a major source of data inconsistency.
Scalable capacity
Higher menu volumes without adding headcount.
Less manual entry
Standard formats now need little to no human intervention from intake to storage.
Better visibility
Automated ticket tracking gives operations real-time insight and makes issues easier to catch early.
