Add a new alert class by text prompt or a handful of image crops, on the box, forward-only, with a gate that checks the new class before it goes live.
Closed source, inquireDhi COVE lets an operator add a detection class on an already-deployed edge box, either by describing it in text or by showing a handful of image crops, without a cloud retrain and without a GPU training loop.
New classes are added forward-only, with no gradient updates to the existing model, and a gate replays the deployment's own evaluation pack before the new class is allowed to go live, catching classes that would degrade what already works.
A real local-LLM path (qwen3.5:2b via Ollama) expands plain-language operator requests into prompt ensembles, with a deterministic offline fallback when the LLM path is unavailable. The class library itself is versioned and rollback-capable.
Toy-provider proof run, seed 7: 4 classes added sequentially, all 4 gates passed, final mAP 1.0 (24 of 24 objects, 0 false positives), average forgetting +0.0000, about 45 ms per class addition including gate replay, 0.31 s total wall time. An adversarial phantom class was correctly rejected by the gate: it would have collapsed existing-class AP to 0.000 with 9.0 false positives per image.
Same demo run with Dhi COVE's real embedding tower (OpenCLIP ViT-B-32) on CPU: all 4 gates passed, new-class AP 1.000, 1.000, 1.000, 0.875, average forgetting +0.0000, 0.36 to 0.53 s per class addition.
Six classes added continually against 454 class-agnostic proposals spanning all 80 COCO categories, RTX GPU. 4 of 6 go live: zero-shot stop_sign AP 1.000, elephant AP 0.563; few-shot zebra AP 0.933, fire_hydrant AP 0.583. Forgetting stayed at 0.0000 across every addition, about 1.4 s per on-box class addition. The gate correctly held back 2 classes the embedding space could not separate few-shot (umbrella AP 0.02).
Today an outsider cannot reproduce any of this: the repo is private, there is no public dataset or demo Space, and there is no LICENSE file governing reuse even if the code were shared. The numbers above are reported from the project's own README and result files, not independently verifiable from outside yet.