01
The context
Education software has several audiences at once. Teachers prepare and review work, pupils use learning materials, and schools need to decide how the technology fits their practices. The infrastructure has to support those distinctions.
My work sits in the systems behind the teacher-facing experience, where limits, permissions and document handling affect whether an apparently simple workflow is dependable.
02
My contribution
I worked on the Marlow agent infrastructure in a production monorepo, with a focus on the components that govern resource use and access.
- Refactoring token rate-limiting behavior.
- Improving authorization isolation around agent workflows.
- Integrating OCR into document-processing pipelines.
- Building deterministic fixtures so behavior can be checked repeatedly.
03
Why these details matter
A rate limiter affects whether an agent can complete its work under load. Authorization determines which information and actions are available in a particular context. OCR sits earlier in the chain: if a document is read incorrectly, later reasoning starts from the wrong input.
Working across these components requires following the whole workflow, then making the individual boundaries easier to test. The fixtures are part of that work, because an improvement needs a reproducible way to establish what changed.
- 01Teaching materials
- 02Document & agent systems
- 03Teacher workflow
04
Connecting engineering and product
Alongside engineering, I study the education-AI competitive landscape and translate findings into workflow and product suggestions. That includes thinking about how AI is introduced to teachers and pupils, and where it supports the task at hand.
This role gives me experience contributing inside a shared codebase while considering the people who use the resulting product.