aionVIS: Agentic Vision Model Training

aionVIS turns one plain-English sentence into a trained, ready-to-deploy computer-vision model. Teaching a camera to recognise something normally means collecting thousands of photos and paying people to draw a box around every object in them. aionVIS removes that need entirely. You say what your model is for, in plain English ("I need to detect toy cars in my house"), and five AI agents do the rest: one designs the training scenes (Gemma 4), one paints them as photorealistic images (FLUX.2), one labels every object it finds (SAM 3), one audits those labels and throws out the bad ones, and the last one trains the model and hands you deployable weights. Nobody collects data. Nobody draws a box. Nobody reviews one. It runs on any GPU, but AMD is what lets the agents work at the same time instead of taking turns: the models need 125 GB of memory, and a single MI300X has 192 GB, so the whole swarm stays loaded on one card while it generates, labels, verifies and trains. The stack is ROCm-native. We initially used Fireworks AI to access Gemma 4, but the language model now runs directly on the AMD GPU—meaning no data and no API calls ever leave the machine. Proof, measured live using AMD Compute (MI300X): 5,000 generated images, a detector trained in 44 minutes for about $1.47 of GPU time, scoring mAP50 0.96. Exported to ONNX, that model now runs on a phone camera and finds real toy cars on a real floor. aionVIS is a real product, live at aionvis.com. Sign in with your own AMD GPU node (endpoint URL plus API key) and the console trains real models on it, end to end. Want to look around first? The same console opens in your browser on simulated data, with no account and no GPU.