Case study · P—03
Contract-Driven Data Platform with GenAI-Assisted Onboarding
- Role
- Architecture and hands-on development · Data Solution Architect
- Period
- June 2026 – ongoing
- Organisation
- LC Waikiki
- Stack
- Python, Pydantic, FastAPI, Playwright, React + TypeScript, Cloud Run, Cloud Storage, BigQuery, Cloud Composer (Airflow), Dataform, Azure OpenAI, Anthropic Claude API
- Status
- Active
In short
New data sources join through validated YAML contracts; LLM agents work at design time only, so production runs stay deterministic.
Context
LC Waikiki is a fashion retailer with 1,000+ stores in 50+ countries plus e-commerce. This cloud-native data platform on Google Cloud uses generative AI for design-time automation and deterministic code in production.
Role: architecture and hands-on development, as Data Solution Architect (since June 2026).
Problem
Adding a data source should not require changes to core code. Production must stay deterministic and auditable, even though LLMs help with onboarding.
Approach
- Contract-driven ingestion: each new source is described in a versioned YAML contract, validated with Pydantic. Core code does not change.
- LLM agents at design time only: a ReAct-style agent with typed reasoning traces, behind a provider-agnostic interface over Azure OpenAI and Anthropic models. Production runs are fully deterministic.
- LLM evaluation: golden datasets; metamorphic, mutation and property-based testing; trajectory grading; prompt-injection defence; regression and drift checks.
- Pipelines as code: Cloud Run Jobs orchestrated by Airflow (Cloud Composer), and BigQuery medallion layers (bronze, silver, gold) in Dataform, generated from metadata.
- Governance: a FastAPI backend (CQRS) and a React/TypeScript admin UI with role-based access control and approval workflows.
Architecture and technology
Python, Pydantic, FastAPI, Playwright, React and TypeScript. On Google Cloud: Cloud Run, Cloud Storage, BigQuery, Cloud Composer (Airflow) and Dataform. Azure OpenAI and the Anthropic Claude API as model providers. Engineering practice: architecture decision records, architecture-boundary tests, automated test suites and AI-assisted development.
Outcome
Work started in June 2026 and continues.
Lessons
- The main design line runs between design time and production. LLM agents work at design time; production runs as deterministic code, so it stays repeatable and auditable.
- The data contract is the interface: a new source is a contract change, not a core code change.