Architecture

You don’t need to know how AcuSight is built to use it — but if you’re curious how a camera on a line becomes part of a live quality system, here’s how the pieces fit.

​

The shape of it

AcuSight has two halves: a fleet of edge devices out on the line, each running a model on-device, and the platform they connect to. The dashboard, API, and Co-Pilot are all views onto that platform. Three sections in this picture are what make the shape unique:

• Every device arrow starts at the fleet. Devices only ever dial out — even “deploy” is the device pulling a new model at its next check-in, never the platform pushing one in. That’s why a camera behind plant NAT needs no open ports.
• Live video skips the platform. It streams peer-to-peer from camera to browser over WebRTC; the Core API only brokers the session, so the server is never in the video path.
• Management is on-demand. The encrypted tunnel for container management exists only while you’re using it, then closes — nothing stays listening on the device.

​

Edge devices run models on-device

Each device runs a lightweight agent plus the model you deploy to it, doing inference on the camera itself — so detection happens at the line, in real time, without round-tripping every frame to a server. Because each device carries its own model, one fleet runs many inspections at once: one camera checking labels, another catching surface defects, a third watching operators for presence and PPE — different models, different lines, a single platform tying them together. Devices keep an outbound connection, so they work behind NAT or a firewall with no inbound ports.

​

How devices connect

Every connection a device makes is outbound — from the device to the platform, never the other way around. That’s why a camera behind plant NAT or a strict firewall needs no port forwarding, no VPN, and no inbound rules:

• Check-in. The edge agent sends a regular heartbeat over HTTPS, authenticated with its per-device key. Each check-in reports status and picks up pending work, such as a newly deployed model.
• Data up. Captured images, detection events, and recording segments travel over the same outbound HTTPS path.
• Management on demand. Container management — the containers, stacks, volumes, and networks you see per device in the dashboard — is powered by Portainer, which AcuSight embeds as its fleet orchestration layer. When you open a device’s containers, the platform signals the agent at its next check-in and the device opens a short-lived encrypted tunnel back to the platform. The tunnel closes again when you’re done; nothing stays listening on the device.

So the platform never needs to reach into your network — devices reach out to it. If a device loses connectivity, it simply resumes checking in when the network returns (see Reliability).

​

The platform

The platform is a set of cooperating services behind a single web app. With the hosted service, AcuSight runs them for you (see Reliability); they can also run on-prem or fully air-gapped. Conceptually they divide the work:

• Core API — the gateway devices talk to: registration, check-in, and brokering the WebRTC session that streams live video peer-to-peer from device to browser.
• Fleet orchestration — real container management across the fleet, built on Portainer: each device’s containers, stacks, volumes, and networks, and group-based deployment of workloads to many devices at once.
• Telemetry — a time-series store (TimescaleDB) that captures every detection and good/defect event, and powers dashboards, OEE, and model-monitoring.
• Datasets & training — projects, batches, annotations, immutable dataset versions, and the training that turns them into models.
• Model registry — versioned models with champion / baseline aliases and the record of which model runs where. See Model lifecycle.
• Object storage — images, dataset versions, model artifacts, and recordings.
• Co-Pilot — a natural-language assistant that reads your live data to answer questions. See Co-Pilot architecture.

Every request is scoped to your organization.

​

How you reach it

Three surfaces sit on the same platform and the same data:

• Dashboard — a live video wall showing every device at once, an events timeline of good/defect/anomaly cards, OEE & quality dashboards, Co-Pilot chat, and device & model ops.
• REST API — the same operations, callable from your own code.
• Co-Pilot — ask about devices, defects, OEE, and models in plain English.

​

How data flows

• Events — on-device inference produces detections and good/defect events that stream to the platform over HTTPS and land in telemetry.
• Live video — streams peer-to-peer from each device to your browser, keeping the video wall real-time and the server out of the video path.
• Metrics — flow to the dashboards as they arrive, so OEE and quality views stay current.
• Models — deploy from the registry back out to the fleet.

​

Where it runs

The same platform runs in two shapes — the services, the agent, and the data model are identical; only where it lives and how you sign in change: Devices behave identically in both — same agent, same outbound-only connectivity. See Air-gapped / on-prem for the offline story.

​

Security

• Users and devices authenticate differently. People sign in through single sign-on (OIDC); devices hold a per-device key issued at provisioning. A compromised device key can’t act as a user, and revoking a device doesn’t touch anyone’s login.
• Everything is organization-scoped. Every API request — including container-management calls down to a single device — is checked against your organization before it goes anywhere.
• Devices expose nothing. No inbound ports, no listening services; management happens over the on-demand encrypted tunnel described above.
• All traffic is TLS, from browser to platform and device to platform alike.

For the full security model — identities, transport, and isolation — see Security & trust.

​

The loop

Put together, the pieces form a loop: devices collect images, you curate them into datasets, the platform trains a model, you deploy it back to devices, and telemetry lets you operate and improve it — then round again.

​

Next steps