Labeltune

Labeltune starts from a precise belief: training models that truly understand Italian requires Italian data, produced by Italian people. Existing data-labeling platforms are built for the global market and treat local language and culture as a detail.

The heart of the product is an annotator community with a mandatory onboarding path of courses and quizzes: nobody touches a dataset before proving they understood the guidelines. On top sits a visual drag-and-drop task builder, with instructions the AI helps enrich.

For audio tasks there is a waveform editor with integrated Whisper transcription, so annotators listen, correct and segment in one place. For clients there is an enterprise panel with a dataset explorer, quality analytics and measurable controls such as inter-annotator agreement (Cohen's Kappa).

I designed the product around the idea of a traceable, reproducible dataset: every example keeps who labeled it, with which instructions and in which version. That is the difference between a pile of labels and a training asset you can trust.

The stack is entirely modern — Next.js 16, React 19, TypeScript, Tailwind CSS 4, Framer Motion and Supabase — and the community counts 4,500+ talents. Labeltune is an attempt to build the data infrastructure that the Italian AI ecosystem is missing.