Unified data platform and AI chatbot for restaurant operations

A restaurant technology startup needed to unify data from multiple sources — EPOS systems, supplier invoices, and market intelligence feeds — into a single analytics platform. Restaurant operators needed to answer questions like "What was my food cost percentage last week?" and "Which suppliers are most expensive for chicken?" without writing SQL or navigating dashboards.

The platform needed near real-time data freshness, the ability to handle messy real-world data from multiple EPOS providers, and a conversational interface accessible to non-technical users.

We designed a 3-layer SQL architecture (staging, canonical, fact) to normalise data from heterogeneous sources into analytics-ready views. AWS Step Functions and Edge Functions provided near real-time ingestion orchestration.

For the conversational interface, we built an agentic chatbot using LangGraph with 50+ Zod-validated tools organised via a registry pattern, giving the AI structured access to every dimension of the business data. The system included budget-aware execution with time limits, tool caps, and graceful degradation, plus LLM-as-a-Judge for automated quality testing. We managed a provider transition from OpenAI to Google Gemini via a provider-agnostic abstraction layer with zero downtime.

The platform went live with restaurant operators able to query their business data conversationally — asking complex multi-step questions that the agent resolves through tool calls, SQL generation, and contextual formatting.

The 3-layer data architecture provides a clean separation between raw ingestion and analytics-ready views, making it straightforward to onboard new data sources and EPOS providers. The provider-agnostic architecture has already enabled one zero-downtime model migration.