What this is for

LabeloxAV is a data engine, not a labelling tool. The distinction matters for how you use it: a labelling tool's job is to get boxes drawn, and this system's job is to decide which boxes are worth a person's attention, then prove whether the ones it drew itself were right.

If you are here to draw boxes on a hundred images, this is heavier than you need.

Use cases it is built for

Turning fleet footage into a training set

The core loop. You ingest dashcam drives, the three-path pipeline labels them, the confidence gate routes everything to auto_accept, review, or annotate, and a person works the review queue in uncertainty order rather than frame order. You export a sealed dataset with a coverage datasheet that states what the release does not know.

Start at Ingest to dataset.

Finding the moments that matter in a long drive

Most frames of a dashcam drive teach a model nothing. SIEVYX exists to find the ones that do: embedding search, near-duplicate suppression, rare-class discovery, and behavioural scenario mining. You can ask for "autorickshaw cutting in at a junction" in natural language and get ranked frames, blended with a rarity term so the answer is not thirty near-identical sedans.

Auditing labels you did not make

Point it at an existing corpus and ask how good the labels are. The judged per-class precision pass (scripts/run_class_precision.py) samples each class, asks a calibrated VLM whether each label is right, and reports precision with the judge's own error corrected out. On this corpus that surfaced a class at 0.02 precision that confidence scores said was fine.

See Audit label quality.

Closing the loop

Train on what you labelled, evaluate per slice rather than in aggregate, and let a champion/challenger gate decide whether the new model ships. A regression on a slice that matters blocks promotion even when the headline metric improved.

Running a second domain on the same engine

The domain is a pack, not the engine. packs/sec serves physical security from the same spine: same ingest, same review, same gate, different ontology and different safety definitions. If your domain is not driving, this is the seam to read: Domain packs.

What it is not

• Not a crowdsourcing platform. There is no task marketplace, no worker payment, no consensus across five anonymous annotators. It assumes a small number of trusted reviewers.
• Not real-time. Nothing here runs on a vehicle. It is an offline engine.
• Not a general-purpose image tool. The ontology, the confusion cliques, the safety definitions and the scene attributes are all shaped around road scenes.

How the pieces fit

Seven platforms, one backend, one flywheel. Each owns a stage and some can block the ones after them.

SANYX      ingest QA          can quarantine a bad session
CALYX      calibration        can block on rig drift
SIEVYX     curation           decides what gets labelled
LABELOX    annotation         auto-label + human review
ORACLYX    pseudo-GT          auto-truths the majority, routes disagreement
VERDYX     evaluation         champion/challenger verdict, can block promotion
FORGYX     edge optimization  quantize and benchmark, can block on latency

Full detail: The seven platforms.