This is not a vague AI concept page. The current codebase already shows an Azure deployment workflow, Supabase-backed auth and storage, OpenAI-powered structured menu extraction, and multiple validation layers between an uploaded PDF and the kitchen workspace.
The current parser is wired to OpenAI and defaults to gpt-5.4. If production moves to a different model, the page should move with it.
This repo proves Azure deployment, but not a specific Azure region. Region and sizing should match the live resource configuration.
Each layer below maps directly to code or workflow already present in this repository. The copy stays close to implemented behavior and avoids inflated claims.
The repo includes an Azure Static Web Apps CI/CD workflow. Every deploy runs lint, typecheck, tests, and a production build before the Azure upload step.
Authentication, relational data, and menu PDF storage are all wired through Supabase. The schema and row-level security migrations live in this codebase.
The parse route calls OpenAI's Responses API and forces the model to return strict JSON matching the menu schema required for the selected venue, dates, services, and item types.
AI extraction is only one layer. The app validates venue alignment, required service slots, item-type matching, signed file access, and parse completeness before data is accepted.
The pipeline is narrow by design: ingest the file, define the allowed schema, run structured extraction, validate hard, then surface the result in the workspace.
A weekly menu PDF is stored in Supabase and tied to a venue and menu period.
The app loads the allowed dates, active services, and item types for that exact venue before generating the extraction schema.
OpenAI receives the PDF plus a schema-constrained prompt and returns machine-readable menu JSON instead of loose prose.
Returned rows are checked against configured services, item types, page numbers, and required service coverage.
Validated items feed the workspace used for menu mapping, production sheets, uploads, recipes, and line setup.
Small details matter more than generic trust badges. These are the kinds of behaviors that make a real product feel legitimate.
The parser creates a five-minute signed URL for each menu PDF before sending it to OpenAI.
The route aborts long-running AI requests instead of hanging indefinitely.
OpenAI output is constrained to a named JSON schema rather than free text.
Coverage is measured against configured date and service requirements.
Culinary Signal is strongest when it says exactly what it does: it turns weekly menus into structured operational data using OpenAI for extraction and application logic for verification, storage, and workflow.
The public site now has a factual system page for buyers, operators, and partners who want to understand what is behind the interface.