Indagate
Industrial agent platform
They read sensors, camera frames, equipment specs and your own SOPs. Every agent is admitted against the device it targets before it can act, and every verdict is retained.
Standards and platforms your team already runs
Each one ships as a pre-built agent you can point at a line and start evaluating, rather than a framework you have to assemble.
Where the cycle time actually went — waiting on material, operator reach, gate cycles, downstream queue. Unsafe states surfaced alongside, beside rated safety control.
ExploreDefects, assembly completeness, labels and date codes — checked against what the work order actually says.
ExploreRestricted entry, unattended items and perimeter activity, turned into events a person can review.
ExploreCondition, drift and anomalies — including analog gauges on assets that produce no data of their own.
ExploreOne agent reading across every record the others wrote, alongside your MES and CMMS.
ExploreDescribe the rule in plain language, or hand it the SOPs and work instructions you already have. Point it at a frame or sensor stream and it returns a structured decision record — the observation, the rule it matched, and what it recommends.
or choose an image from this device
Frame content is sent directly to Bedrock for this request and is not stored by this demo.
This browser demo runs the model in the cloud so you can try it without hardware. Deployed agents run the same model on your device.
The observation, matched rule and recommended operator action will appear here.
Use a conveyor, robot cell or loading area
! Scenario and PLC action are simulated. Image analysis uses a real model on AWS; no equipment or PLC is connected.
The visual evidence, model recommendation and simulated PLC record will appear here.
Zone air quality + AHU response · hourly
Loading the Indagate reference building…
GM102B Raw ends at 313 raw. Fan feedback remains between 70.73% and 73.46%, with an estimated final speed of 1,314 rpm; AHU condition is stable within the supplied sample.
Air-quality and AHU samples were independently sourced and aligned by observation index, so their overlap is not evidence of causation. RPM is estimated from measured VFD percent using an assumed 1,800-rpm rating. Live refers to the demo and Argo analysis—not a live stream.
Sanitized CSV in · live Argo analysis out
It holds your hardware spec, your operating rules, and the decision it just made. So “why did line 3 stop” gets one answer with the evidence attached — instead of a reconstruction across four consoles and a phone call.
Argo provides decision support. Human review and safety-rated controls remain authoritative.
Most platforms give you one of these. An industrial deployment needs all three.
Perception, reasoning and tool calls in one loop. It holds what it saw, works against a rule you wrote in plain language, decides what to do next, and can be questioned about any of it afterwards.
See it runIndustrial systems will not take an agent on trust. Each one arrives with a publisher, a version, the device it targets and the reach you granted it — all checked at the gate, so the security review happens before deployment rather than after an incident.
The Oddy testWe come from embedded and industrial hardware. Send us your installed base and we will qualify what it can actually run, model by model. Where it genuinely falls short, our reference controller ships as a drop-in.
Supported targetsThree checks. One named device. A verdict you can produce two years later.
In 1973 British Museum conservation scientist Andrew Oddy published a test that exposed silver, copper and lead coupons to a candidate material before it was allowed near an irreplaceable object. You do not test the object. You test the thing you are about to introduce, under the conditions it will actually meet. An agent is a candidate material. Your PLC-connected line is the object.
How the three coupons workTarget-scoped software admission
CANDIDATEqwen2.5-vl-3b · gguf q4_k_m
TARGETQCS6490 · line 3
Publisher, version and SHA-256 digest match the signed manifest.
Memory, runtime and accelerator sit inside the device envelope.
Requests nothing beyond the reach you granted on that device.
A candidate arrives with its publisher, version, digest, intended role, requested reach and target.
Three coupons run against the specific device the candidate wants to reach.
A pass may install to a versioned slot. A fail never receives an edge directive.
The evidence and verdict stay on record so you can answer what ran and why.
We come from embedded and industrial hardware, so this is a question we can answer precisely. Send us your installed base and we will qualify what each device can run. Where existing compute falls short, our QCS6490 reference controller ships as a drop-in. Either path keeps the same deployment contract: a manifest-scoped runtime and a bounded field edge.
Inference runs on the device. Frames are processed, models are executed, and language models run at the edge — so video never leaves the building and there is no per-token meter on your production line. Preprocessing and anonymization happen on the device too, which means what does leave is already stripped of what you did not want to send.
Fleet management, model distribution and long-term dataset storage run on AWS, because managing a hundred devices and retaining two years of evidence are jobs a control plane does better than a controller bolted to a machine.
The boundary is explicit and inspectable. Every agent’s manifest states where it executes and what it may reach. You can read that before it installs, and read it again two years later.
The usual trade is a system IT will approve and the floor resents, or one the floor likes and IT blocks. Governance here is part of how the agents run rather than a layer added afterwards — so IT and OT are signing off on the same system, not negotiating over it.
Start on one line with one defect or event. Prove the signal exists before anyone talks about a rollout, and keep the operator in the loop throughout.
Fits the plant you already run. Telemetry crosses a bounded companion MCU; control authority never moves.
Every agent is signed, scoped and recorded before install. Reach is declared up front, execution location is stated, and the evidence is retained.
Agents that touch operational technology get assessed against real regimes. Indagate is built so the evidence those regimes ask for is a by-product of running the system, not a project you start once someone requests it.
What the admission record maps to, and how far the claim goes:
Support is not certification. Indagate holds no certification against any of these eight frameworks and none is implied. These are the obligations the system is designed to produce evidence for; conformity remains the operator’s assessment, made with your own auditors against your own deployment.
Send your SOPs and we will audit them free — the agent that fits your procedures, and the hardware it runs on.
Read how admission and the runtime actually work, or run the CLI against a device you already have.