Case study
MEF — Metadata Execution Framework
NatWest (via Accenture) · Oct 2022 – Jun 2026 · Principal Engineer / Platform Owner
A declarative platform for regulated data delivery — one workflow model, generated Airflow/Spark artifacts, and governance enforced in CI.
Replaced team-specific workflow implementations with one governed platform model for delivery across multiple bank teams.
The problem
Many delivery teams were building data workflows independently: inconsistent Airflow and Spark implementations, duplicated patterns, governance and operational standards varying team by team. Hard to onboard new teams; hard to audit what ran where — in an environment where auditability is not optional.
The platform
I designed and led MEF as a declarative execution layer: teams model workflows in a DSL, while the platform validates them, resolves dependencies, packages and publishes releases, and generates runtime artifacts instead of relying on hand-written pipeline code.
I also exposed the platform through a FastAPI metadata API and an MCP layer, so AI assistants could discover and invoke governed platform capabilities instead of relying on ad-hoc integrations.
I drove adoption through workshops and cross-team enablement, and supported the 2026 handover to bank infrastructure ownership.
Outcomes
- ▪Designed and led the declarative execution framework used as the shared delivery model across multiple bank teams
- ▪Replaced hand-written Airflow DAGs, Spark configs, and deployment descriptors with artifacts generated from one declarative workflow model
- ▪Moved governance into CI: validation, standards, auditability, and governed AI access via MCP
- ▪Owned the platform from concept to adoption and supported the 2026 handover to bank infrastructure ownership
Why it matters for AI product work
MEF's metadata already described operators, parameters, and execution semantics — so exposing capabilities to AI assistants through a metadata API and MCP was a natural extension: governed discovery instead of every team wiring its own integrations. Intent (what) is separated from runtime (how) and infrastructure (where). The same pattern extends to AI systems: declarative capabilities, approved tools and models, generated configurations, and governance built in — whether the runtime is Airflow or an agent loop.