I build systems that survive contact with production.

My work centres on systems where the backend carries real business rules rather than shuttling database rows — billing and subscription lifecycles, multi-tenancy, background processing, audit trails, and integrations that fail in interesting ways.

I design for the second year of a codebase, not the first sprint. Shared contracts so an API change breaks the build instead of production. Composable feature bundles so the tenth feature costs less than the third. Cross-cutting concerns built once as infrastructure, not re-implemented per feature.

And I stay responsible after the deploy — containers, certificates, DNS, database migrations against live data, and the failures that look like application bugs until you trace them down a layer.

Platform foundation

One contracts-first monorepo, reused as the starting point for every new product.

Five packages share a single type system. A contracts package holds every domain's types, enums and schemas; the API and the client both compile against it. Break an API shape and the build fails instead of production. Every other decision hangs off that one.

What it removes from day one of a new product: wiring up deployment, environments, auth, tenancy, permissions and access control. Those are solved in the foundation, which is why it became the shape the whole portfolio is built on rather than each product starting from scratch.

Security

Defence in depth, built as shared infrastructure.

Security lives in the platform layer rather than in individual features, so it applies by default and every exception is an explicit, greppable opt-out on a single route. Most codebases have two or three of these layers; the value is in having the set apply consistently.

AI engineering

Agent tooling I've shipped, not just prompted.

I built a Model Context Protocol server as a first-class package inside a production platform — dual HTTP and stdio transports, a tool registry, and typed tools, resources and prompts that reach the API through an authenticated client. It gives AI agents a governed interface to the product instead of raw database access.

Day to day I work with Claude, Claude Code and Cursor for implementation, architecture exploration and debugging unfamiliar systems, with repo conventions written as agent-readable rules so humans and AI converge on the same patterns. n8n for AI-backed workflow automation.

The useful framing: conventional engineering judgement, accelerated — from problem definition to working implementation in less time, without giving up on knowing why it works.

Exploration

Procedural character animation, in Unity.

An ongoing side build for a cricket simulation, where the character's body adapts to its actual environment instead of replaying fixed clips — layered inverse kinematics for foot placement and ground adaptation, hand targets, head aim, and driven hip, chest and shoulder adjustment.

It means working directly in quaternion rotation, transform-space conversion and runtime interpolation to keep joint chains stable through movement, jumping and diving. I keep it going because the hard problems there are nothing like the ones at work.