Make your data
trustworthy.
One workspace to annotate, validate, review and release high-quality datasets — with evidence at every step.
Recent datasets
Annotation workflow
From raw files to reliable releases
Confidence is not
the same as truth.
THEO preserves uncertainty, evidence and human decisions. Automated proposals never become ground truth by default.
Datasets
Organise raw data, run annotation providers and manage releases.
Review queue
Inspect every proposal. Approve, correct or reject with a traceable decision.
Quality & releases
Quality signals, explicit limitations and immutable dataset snapshots.
Evaluation lab
Evaluate predictions against independently verified references. Metrics stay separate; no synthetic trust score.
Classification evaluation
One label per line, aligned by sample.
Confidence calibration
Pair confidence and correctness by sample.
Evaluation output
Record dataset version, sampling method and gold-label provenance alongside these results.
Choose an evaluation to see results.
Calibration requires representative, independently verified labels. Agreement metrics measure annotator consistency, not correctness. Subgroup audits require explicit, ethically collected group data and expert interpretation.
Activity log
A chronological record of important dataset and annotation events.
Built with intent.
A modular foundation designed for reproducibility, extensibility and honest quality reporting.
Dataset control plane
Dataset registry, metadata, file ingestion, version manifests and API contracts.
Provider layer
Six modality pipelines with optional pretrained inference adapters, explicit model references and human review gates.
Review & quality plane
Calibration, classification metrics, annotator agreement, subgroup disparity signals, adjudication and immutable releases.
Operator workspace
Responsive glass interface with dataset management, review queue and exports.
Model-backed inference is opt-in and depends on configured checkpoints, optional dependencies, hardware and model licences. Current pipelines include YOLO detection/tracking, Transformers classification/NER, optional Tesseract OCR, faster-whisper transcription and pyannote diarization. No generic activity-recognition or universal segmentation guarantee is made. Calibration and fairness reports require independent gold data and explicit subgroup labels. Outputs remain proposals until human review.