Frequently Asked Questions

Images, video, 3D point clouds, LiDAR, radar, audio, text, documents, time-series and multi-sensor fused datasets. If your data does not fit those descriptions, tell us what it is — the interesting projects usually do not.

Against a gold-standard set agreed with you before production, using inter-annotator agreement and sampled independent audit. Thresholds are written into the engagement; a batch below threshold is reworked at our cost, not delivered with an apology.

Yes. We can work inside your platform, deliver into it through its API, or run the programme on ours and export in your format. The tool is rarely the deciding factor.

A pilot batch typically begins within a week of agreeing the guidelines. Scaling to volume follows once the pilot's agreement scores hold. We do not start volume work on an unproven guideline.

Role-based access, project isolation, encryption in transit and at rest, and de-identification before annotation wherever the task permits. Restricted programmes run on controlled environments with no download rights. Terms are written into a data processing agreement.

COCO, YOLO, Pascal VOC, KITTI, JSON, CSV and custom schemas. Delivery can land directly in your cloud storage or training pipeline rather than as a download.

For pre-labelling, yes — and every pre-label is reviewed by a person before it ships. We measure the correction rate and run blind unassisted control batches so the effect on quality is a number rather than a promise.

Text and speech work is delivered across major European and Indian languages, with others on request. Tell us the language and the domain; the domain is usually the harder half.

By unit of work — per image, per frame, per document, per hour — depending on the task, with volume tiers. A pilot gives both sides a real number before anyone commits to scale.

Agreement metrics are monitored continuously, so it surfaces during production rather than at delivery. The batch is reworked, the cause is investigated, and if it is an ambiguous guideline — which it usually is — the guideline is fixed and the affected data identified.