Mabyduck
A platform for running perceptual audio experiments at research scale
Intro
Mabyduck is an AI research platform for perceptual audio evaluation. Teams upload a dataset, run listening experiments, MUSHRA and pairwise comparison, against a pool of testers, and turn the results into a report. We designed the platform end to end: the experiment wizard, the dataset and results models, the reporting, and the account and rubric systems that hold it all together.
Onboarding
First contact with the platform. We designed a sign-up and onboarding flow that gets researchers into their first experiment quickly, without asking them to understand the whole system before they've run anything.
Experiments
The heart of the product. We designed the full experiment lifecycle, from the wizard that walks a researcher through setup, to choosing a test type like MUSHRA or pairwise comparison, to the live detail view that tracks a run as results come in. The goal was to make a rigorous research method feel like a few clear steps.
Datasets
Every experiment starts with audio. We designed the dataset model and its upload flow so researchers can bring in their files, organise them, and reuse them across runs, the raw material the rest of the platform depends on.
Reports
Where a run becomes a result. We designed the reporting experience to turn listening data into clear, shareable findings, so a team can read what the testers heard and act on it without wrangling the numbers themselves.
Rubrics
The scoring logic behind every experiment. We designed the rubric system that defines how testers rate what they hear, giving researchers control over the criteria while keeping results consistent from one run to the next.
Search and Notifications
The connective tissue of the platform. We designed search and notifications to keep researchers oriented across experiments, datasets, and reports, and to surface what's changed the moment it matters.
Settings
Account, team, and billing in one predictable place. We designed the settings experience to be clear and low-friction, the kind of admin surface people trust because it never surprises them.
