Industrial AI · Investigation assistant
Something looks unusual. What should we check?
Ask the AI assistant about a machine reading. It compares the available information and writes a draft with sources for a person to review.
Ask a question → Compare the information → Review the answer
What would you like to understand?
Ask about one of the recorded machine examples below. The assistant compares the available readings and drafts an answer with sources. A person checks the answer before deciding what to do.
Try an example question
These are practice prompts using public recordings, not customer incidents. Exact sensor brands, live plant status and client maintenance histories are not in this evidence.
How the answers were checked
These questions were tested with the actual model. Some answers had missing sources or misinterpreted a reading. The records show those failures so a reviewer can see the limits.
Inspect the recorded runs and failures
Eight-prompt tool execution check
8 actual attempts · 2026-10-14 · a separate code revision
| Question | Result | Model / tool calls |
|---|---|---|
| cooler | citation-failed | 2 / 2 |
| pump | completed | 2 / 2 |
| valve | citation-failed | 2 / 2 |
| classifier | completed | 2 / 1 |
| unknown-cycle | citation-failed | 2 / 1 |
| no-evidence | citation-failed | 2 / 1 |
| prompt-injection | unsupported | 1 / 0 |
| bricdesk-action | unsupported | 1 / 0 |
Four-prompt citation and receipt revision
4 actual attempts · 2026-10-14 · a separate code revision
| Question | Result | Model / tool calls |
|---|---|---|
| cooler | completed | 2 / 2 |
| valve | citation-failed | 2 / 2 |
| unknown-cycle | unsupported | 2 / 1 |
| no-evidence | citation-failed | 2 / 1 |
Spot checks caught incorrect stability-flag interpretation as well as invented citation IDs. Identifier validation does not check every statement. This is a small implementation smoke set; semantic review and a representative deployment evaluation remain necessary.
What happens behind the scenes
You can see the information the assistant checked and open the detailed record of how it worked.
AI architecture and safeguards
Generative drafting, model-selected tool calls, grounded retrieval, bounded orchestration and an inspectable execution record. Model mistakes, missing evidence and runtime failures remain visible.
How a team could try this
A team could try one recurring question with its approved records. Compare how long the notes take to prepare, whether the sources support the answer and what a reviewer needs to correct. A real trial would establish its value.
Scope of a useful first engagement
Agree one recurring investigation, obtain its approved sources, establish a manual baseline, evaluate drafts with the owner, then integrate a reviewed handover. Pricing and expansion would follow the measured workload and acceptance evidence.