QUARKEONLABS / AI ENGINEERING
← All projects

Vision / Evaluated demonstration

Label Foundry

From a prompt to a dataset you can review.

Real demo recording · English narration and captions · Fictional data

THE PROBLEM

Annotation work needs faster proposals while retaining control of labels, boundaries and quality.

THE WORKFLOW

Local SAM3 proposes six bottle masks in an original scene. The annotator reviews boundaries, corrects boxes and labels, and exports PNG masks, COCO with exact RLE or YOLO boxes with the image.

  1. 01Image and prompt
  2. 02Local instance masks
  3. 03Annotator review
  4. 04COCO, YOLO and PNG export

RECORDED EVIDENCE

Results, with context.

Results on synthetic data
CheckCases
Proposal count on the scene1/1
Generated masks (structural check)6/6
Dataset
original-controlled-fixtures-v1
Completed cases
1/1
Median / p95 per case
8.433 s / 8.433 s
Evaluation record (UTC)
2026-10-08

Source: the project’s evaluation record, using local inference. Recorded hardware: NVIDIA GeForce RTX 4070 Laptop GPU / 8 GB. Latency includes the full HTTP case and may include initial model loading. These measurements have not been rerun on the machine serving this website.

Case study (PDF · EN) ↓Evaluation data (JSON) ↓Video (MP4) ↓

WHAT WE COULD BUILD WITH YOU

An assisted annotation tool with human control, provenance and exports for your training format.

What we would define first

Taxonomy, annotation guidelines, output formats, quality review and independent training/validation splits.

Demonstration limits

One original AI-generated scene. Six mask outputs and their exports are checked without claiming segmentation IoU. Manual boxes are identified as human annotations. The single-image archive is not a validated training dataset.

Discuss your use case ↗
Screenshot of the interface for Label Foundry
Demonstration review interface.

Explore other projects