What expert labelling covers
Labelling turns raw inputs into data a model can learn from. Saolabs labels images, video, audio and sensor data, from bounding boxes and segmentation to event tagging, transcription and classification.
The difference is who does the labelling. A radiology image is labelled by someone trained to read one. A recording of a machine fault is tagged by an engineer who knows what that fault sounds like.
- Images: classification, bounding boxes, segmentation, attributes
- Video: object tracking, event and action tagging, temporal segments
- Audio: transcription, speaker and event labels, quality judgements
- Sensor data: event labels and annotations on robotics and physical-world logs
Why precision depends on expertise
Many labelling tasks look simple until you reach the edge cases. Is that shadow a lesion or an artefact? Is this vibration trace a fault or normal start-up behaviour? A general annotator will pick an answer. A specialist will often pick a different one, and can say why.
Those edge cases are where models fail in production. Labelling them correctly, or marking them as genuinely uncertain, matters more than raising throughput on the easy items.
Guidelines that experts can actually apply
Good labels start with good guidelines. We write them with your team, test them on a sample set and revise them when experts hit cases the first draft did not anticipate.
We treat disagreement between labellers as a signal. When two qualified experts label the same item differently, that usually points to an ambiguous guideline or a genuinely hard case, and both are worth knowing about before you train on the data.
Sensor, robotics and physical-world data
Models that act in the physical world learn from sensor logs, simulation runs and multimodal recordings. Labelling these needs people who understand the system that produced the data, not just the format it arrives in.
Saolabs matches engineers and scientists to this work. They label events, failure modes and states in the data, and note where the signal alone is not enough to decide.
How a labelling project runs
We scope the task with you: label taxonomy, edge cases, formats and what the data is for. We then label a pilot set so your team can check quality and push back on the guidelines before production starts.
In production, matched experts label the data, with expert review and quality checks throughout. Delivery includes the labelled data in your agreed format and notes on known limits, such as classes with few examples or items experts could not resolve.