An AI-native system-of-record for optics and photonics research and development design automation
TL;DR
It is argued that an optical laboratory reduces to two primitives, the components on the bench and the connections between them, and that promoting this decomposition to a first-class, versioned graph, rather than a file hierarchy, is the substrate that makes both human retrieval and AI-assisted design automation tractable.
Abstract
Documentation in optics and photonics research and development is overwhelmingly unstructured (spreadsheets, plots, instrument exports, images, and free-text notes) and is scattered across shared drives and nested folder hierarchies. A folder tree imposes a single rigid access path, yet any one artifact is relevant along many dimensions at once (by wavelength, by component, by project, by person, by team, or by date), so the tree is the wrong index for almost every question and gives an automated agent no relational signal, only blind path-guessing. We argue that an optical laboratory reduces to two primitives, the components on the bench and the connections between them (with the convenience of grouping a reusable assembly into a modular component that can be saved and reused). Promoting this decomposition to a first-class, versioned graph, rather than a file hierarchy, is the substrate that makes both human retrieval and AI-assisted design automation tractable. We present an AI-native system-of-record built on this principle. It treats a hardware configuration much as version control treats source code: a user can branch a setup to explore an option or debug a problem without disturbing the main line, and each configuration is tied to the specific measurement, parameter, or test it was built for. Every change, by a scientist or by the AI, is captured in an append-only audit log, so nothing is ever truly lost: an accidental deletion or edit can be rolled back, and the whole development history of an experiment, setup, or project stays visible, what was tried, which decisions were made, and where problems arose. Unlike software, where an AI has the whole codebase and the internet for context, hardware leaves it blind to everything outside the computer; a components-and-connections graph gives the AI the same structured view of the bench that a human has, and the AI in turn can fold the free-form artifacts scientists already produce into that structure, acting over it through the same interface as human users. We have implemented the full architecture as a working prototype and report its design and our experience building it.
