LinkedAI

Context

A platform to create and curate training datasets for AI and computer vision models, where the quality of the annotation directly determines the quality of the model.

Problem

Annotating and reviewing thousands of images requires precise graphic tools that stay responsive — a generic UI is not enough.

Constraints

• Large data volumes: projects, datasets, images and review states, all interconnected.
• Drawing and editing directly over images, where precision and responsiveness are the product.

The work

• Built interactive tools for annotation, segmentation and object editing on top of images using HTML Canvas.
• Handled complex application state across projects, datasets, images and review flows.
• Contributed to AI-assisted labeling features and to unit and integration testing.

Decisions

• Canvas over DOM for the annotation surface: the only way to keep drawing precise and smooth at scale.
• Normalized state so that images, annotations and review status could evolve independently.

Result

Graphic tooling used in the day-to-day annotation and review flow of datasets that feed computer vision models.

Learnings

• This is where frontend stopped being 'screens' for me and became tooling with real engineering constraints.
• My first close contact with how AI products actually get built, from the data side.