Aside from using AI in my own work, I focus on making AI something developers can use safely and productively. Here are a few brief examples of how I approach it, in broad strokes.
The problem. Turning a ticket into something a developer can start on takes a lot of context-gathering before any code is written.
The approach. Agentic flows that take a DevOps ticket through to a structured code plan, with a developer reviewing and steering at each stage.
The outcome. Developers stay in control, but the admin-brden, requirements documenting and context injection are all done with greater ease; allowing them to get on with the creative bits they enjoy.
The problem. AI assistants will happily ignore security and compliance conventions unless they are told, every time, what they are.
The approach. Developing a collection of shared, in-house skills and agent rules that carry those conventions with the tool, so safe behaviour is the default and not something each developer has to remember.
The outcome. Tailored in-house rules to help with various areas of development from boilerplating to code-reviews, repo analysis to CVE vulnerability handling.
The problem. Product and company knowledge is scattered, and code is hard to reason about from text search alone.
The approach. Retrieval-augmented generation combined with abstract syntax tree analysis, so answers draw on both documentation and the actual structure of the code.
The outcome. An active, growing knowledge-base for developers and product owners that combines both documentation and code structure, making it easier to reason about and work with complex codebases.
The problem. Some work needs predictable output and tight control over where data goes.
The approach. Running models locally where it matters, wrapped in guardrails for code quality, documentation and framework-upgrade compatibility.
The outcome. This is an on-going research project, including benchmarking and analysis of model efficiency and reliability within controlled environments.