AI Engineering Intern

About the role

At Light, we’re redefining how software is built, how it’s delivered, and how it works.

We’re building organic software. Software that heals itself. Software that adapts to the user and the business. Software that learns and gets better as it’s used.

Our starting point is finance. Companies including Legora, Lovable, and Fuse Energy already run on Light. We’re building towards a platform that understands how each business operates and evolves with it.

As an AI Engineering Intern, you’ll help make that happen: building systems that identify mistakes, improve their own behaviour, and prove those improvements work. You’ll work on the agents, models, and infrastructure that turn organic software into something customers can rely on.

Who this is for

You must meet both of these requirements:

- Currently studying at University of Oxford, or University of Cambridge.

- A silver medal or higher at the International Mathematical Olympiad (IMO) or International Olympiad in Informatics (IOI), or an equivalent level of achievement in another highly competitive endeavour.

If you’re applying with an equivalent achievement, explain the competition or field, your result, and the standard required to achieve it. If you’re at Oxford/Cambridge and you don’t have an equivalent level of achievement we encourage you to not apply.

Our bar for internships at Light is high. You will need the curiosity to explore unfamiliar areas, the willingness to go the extra mile to get things right, the discipline to test your assumptions, and the persistence to keep going when your first approach does not work.

What you’ll work on

You’ll have access to our repositories, real engineering problems, and an internal coding agent that already opens pull requests against production code. Your work will help make these systems more capable and reliable.

- Help agents learn from their work. Our agent reviews merged pull requests and corrections to update its repository-specific instructions. You’ll shape that process, catch misleading lessons, and test whether new instructions improve its performance.

- Train and deploy LLM and ML models. Fine-tune SoA LLM models. Run experiments, compare results against the existing pipeline, and establish whether a model is ready for production. Leverage ML tools to enhance the ability of our Agents via e.g. K-means, t-sne, PCA, random forest trees etc. etc.

- Build Agents. Take full end to end ownership of an agent. Agent do pre-emptive support for the self-healing pipeline; they fix; they communicate with customers; etc.

- Build evaluations and review tools. Define what a correct answer looks like, create tests that expose failures, and help catch poor output before it reaches customers.

- Own projects through delivery. Scope the problem, build the solution, debug regressions, and follow through until the customer can rely on what you’ve shipped.