The dataset-to-model workspace
for visual AI.
Anastig Studio brings together datasets, annotation, preprocessing, augmentation, export, and model orchestration in one workspace. Annotation is only one module.
Everything from image to inference.
Each module is independent. Start with annotation and export; connect training workers and deployment later.
Annotation Studio
Bounding boxes, classification, keypoints, and polygon annotation.
Open moduleTraining
Connect GPU workers and launch model training jobs.
On the roadmapDeployment
Deploy to APIs, edge devices, Docker, and embedded hardware.
On the roadmapEvaluation
mAP, precision/recall, confusion matrices, and model comparison.
On the roadmapWorkflows
Build multi-step visual AI pipelines with a visual canvas.
On the roadmapLabel faster. Review everything.
Anastig speeds up labeling with AI you stay in control of. Every suggestion is a reviewable proposal — you accept, edit, or reject it. Nothing is labeled silently.
Smart Select (on-device)
Draw a box and get a tight object mask from SAM-family models that run in your browser — images never leave your device for this step.
Visual Similarity
Label one example, then find more instances of the same object across your images — one-example assistance that cuts repetitive work.
Reviewable proposals
AI suggests; you decide. Every proposal is accepted, edited, or rejected by a human — a deliberate human-in-the-loop step that keeps datasets clean.
On-device models install once, with per-file checksum verification, before first use. Heavier detection and segmentation run in Anastig Vision Lab on a GPU worker your organization controls — so your images stay in your boundary.
Studio feature matrix.
Honest statuses: Live is in production, Beta is usable but maturing, Roadmap is planned.
| Capability | What it does | Who benefits | Status |
|---|---|---|---|
| Browser annotation (box + polygon) | Draw and edit labels on a fast canvas, fully in the browser, with undo/redo and review states. | Annotators, domain experts | Live |
| On-device AI segmentation | Box-to-mask with SAM-family models that run on your device — no image upload for inference. | Annotators | Live |
| Visual Similarity (one example) | Find more instances of the same object from a single labeled example. | Annotators | Beta |
| Custom labels & negative samples | Define your own classes and manage defect-free samples for balanced training. | ML engineers | Live |
| Dataset versions & health checks | Version datasets and catch class imbalance and train/val/test leakage. | ML engineers | Live |
| Validated multi-format export | YOLO, YOLO-seg, COCO, Pascal VOC, JSON, and classification CSV — checked before download. | ML engineers | Live |
| Preprocessing & augmentation | Resize, crop, tile, plus flips, rotations, and noise — geometry-aware at export time. | ML engineers | Live |
| Model training & deployment | Connect a GPU worker your org controls; hosted training and deployment are planned. | Teams | Roadmap |
Built to stay out of your way.
Canvas work stays in the browser
Annotation, export, validation, and label operations run entirely in your browser. No server round-trips while you draw.
GPU stays external
Training and inference never run on the web server. Connect your own GPU workers or use cloud workers when ready.
Data stays in your boundary
Images stay in your browser in demo mode. In production, your data lives in your tenant and is never used to train shared models.
Start with annotation. Scale from there.
Open the demo to annotate your own images now, or request a pilot to scope a complete dataset-to-model workflow.