Vision AI model creation

Infrastructure monitoring has always relied on numeric sensor data — flow, pressure, temperature, level. But many of the conditions that matter most are best assessed visually: is the inlet grate blocked? Is the pipe corroded? Has the water reached a critical level on the staff gauge? Is the access cover ajar?

Waltero Vision AI closes that gap. Vision-capable W-Sensors send still images from the asset, and you train a custom model in Mimir that answers your inspection question automatically — replacing manual site visits with a model that runs every time a fresh image arrives.

What it is (and isn’t)

Vision AI is for visual inspections of infrastructure conditions captured by your sensors. It is not a generic image-recognition service, and it is not the on-device digit recognition that reads water and energy meters — that’s a separate edge model that ships in the W-Sensor itself.

Typical Vision AI questions look like:

• Yes/No. “Is the screen blocked?” “Is the chamber flooded?” “Is the lid in place?”
• Categories. “Pipe condition: clean / fouled / corroded / failing.”
• Scale. “Blockage level 0–100%.” “Water level 0–10 on the staff gauge.”

If the answer is something a trained person could give by looking at an image of your asset, Vision AI can usually be trained to give it too.

Where it lives in Mimir

Two places:

• Vision Workspace (/dashboard/vision) — an image-analysis workspace for vision-capable sensors. Scope a date range and an organization, browse the captured images, and pipe them through deployed AI models for automated interpretation.
• AI Models — the catalogue of your models, with their current status (Draft, Ready, Deployed) and a one-click deploy.

How a model gets built

Vision AI is no-code. The five-step wizard in Mimir walks you from a blank model to a deployed automation in a single sitting — no ML team, no data pipeline to maintain.

1. Define Your Model

Name it, write a short description of what the AI should look for, and pick the analysis type — Yes/No, Categories, or Scale. The analysis type fixes the shape of the output and the question you’re asking.

2. Provide Reference Images

Open the Device Image Browser, select 3 to 10 reference images from your sensor fleet, and annotate each one with the correct answer. The model learns the visual pattern that maps an image to the correct interpretation directly from your examples.

3. Select Provider & Review Prompt

Mimir auto-generates a system prompt that captures your configuration and reference annotations. Most teams ship the auto-generated prompt as-is; advanced users can edit it directly for tone, edge cases, or output structure.

4. Test Your Model

Run the model against 1–5 unseen test images before you ever deploy. The wizard shows the model’s prediction next to your ground truth so you can spot disagreements early. Iterate on the references or the prompt until the model is solid on real-world inputs.

5. Review & Deploy

Save as Draft, mark Ready for review, or Approve & Deploy straight to production. Deployment also creates the device-level automation that runs the model against every new image — flip it on per sensor from the device’s Automations tab.

Model lifecycle

What you get out of the box

• No-Code AI — the whole flow lives in the Mimir UI. No data pipeline, no Python, no model registry to manage.
• Your data, your model — trained on images from your own sensors, so the model understands your specific assets.
• Editable prompts — full control of the system prompt, from auto-generated defaults to fully custom instructions.
• Continuous testing — validate against real images before deploying, and re-test whenever you update references.
• Seamless automation — deploying a model auto-creates the device-level automation, with per-device activation control.
• Permission-gated — full RBAC integration ensures only authorised users can create, test, deploy, or modify models.

Where the predictions go

Once a model is deployed and activated on a sensor, every new image it captures runs through the model. The prediction lands in the device’s Data tab alongside the rest of its telemetry, so you can trend it, alarm on it, and feed it into the same automations engine you already use for numeric measures.

Costs and limits

AI inference is billable per processed image. The activation banner on each device makes the expected cost visible before you commit. AI Vision tiers are subscription-gated — talk to your Waltero account contact if you need to unlock a higher tier or run a large pilot.

Talk to us

If you have a visual-inspection question you’d like to automate — and a handful of images that show the answer — we’ll help you scope a model and set up the sensors to capture the right images.