Arato Studio
Arato Studio was Arato's environment for testing and comparing LLM prompts, models, and context strategies before shipping anything to production — compare variations side-by-side, run evals, track versioned experiment history, and let product, engineering, and legal all contribute without needing to code.
I worked on Studio from September 2024 to November 2025, as a product designer working closely alongside Ori, Studio's product design lead. We worked more like close collaborators than a strict lead-and-report structure — remixing each other's work, auditing and critiquing when it mattered, some features fully shared and some owned solo, but almost everything touched both of us in some way.
One of the core problems we had to solve from the start: almost nobody who'd use Studio already knew how to work with prompts, variables, parameters, datasets, or evals. There was no existing mental model to build on — we had to invent one, together, from the first sketch.
Ori thought through the earliest version of that mental model — what it could feel like before it had a name. From there, we designed it together: how do you make something this technical feel approachable? It needed to be friendly, understandable, a little playful even — while still being genuinely usable for people doing real, high-stakes work.
Alongside the shared work, I owned a few things end to end:
- Model Migration — testing new models against existing workflows before switching, so teams could migrate for optimization or model deprecation with consistency assured and no production surprises.
- Statistical Confidence — bringing statistical rigor to how run results were compared, so a "better" result was backed by more than a gut read.
- Arato Suggest — AI-powered recommendations to optimize prompts, evaluations, and datasets based on test results.
- Design system work supporting Studio more broadly.
