**Background**

When physical equipment misbehaves, its users face an uncomfortable choice: call an expert, or muddle through on their own. Calling an expert is expensive and slow — scheduling, travel and minimum call-out fees add up even when the fault turns out to be trivial. Muddling through risks making things worse, and can cross lines that void warranties or insurance ("open that screw and the warranty is gone"). Between these two options sits a middle ground: a guided diagnostic companion that gives users enough structured support to resolve straightforward issues themselves, stay safely inside the boundaries of what they are allowed to touch, and escalate intelligently when the problem is beyond their reach. That an LLM can extract troubleshooting steps from an unstructured manual is no longer the question — of course it can. The question is *repeatable, constant quality*: the same fault must produce the same verified guidance every time. A promising pattern: the system drafts step-by-step procedures from the manual, an asset expert verifies the draft, and the verified draft becomes a fixed instruction set the AI then follows — combining LLM flexibility with expert-controlled reliability. On top of the official manual sits a second layer worth capturing: the experiential knowledge of seasoned technicians ("give it a light knock before opening that screw, it comes loose easier"). To keep the focus on this hard problem rather than on learning how industrial machinery works, the project is deliberately set on a familiar class of physical assets — home appliances such as dishwashers, washing machines or AC units— as a stand-in for industrial equipment. The mechanics transfer directly to industrial field service.

**Problem** Key questions to be answered in the project:
- How can an LLM-based guide produce repeatable, constant-quality troubleshooting from unstructured manuals — same fault, same steps, every time?
- What does the expert-verification workflow look like in practice: how are drafted procedures reviewed, locked as instruction sets, and kept up to date?
- How should the interface communicate safe boundaries — what the user may and may not touch, where warranty and insurance limits run — without frustrating the user, while actively preventing harmful interventions?
- How can experiential technician knowledge be captured and layered on top of official manuals?
- What happens when the user gets stuck mid-procedure (step 4 fails)? How does the system genuinely take that feedback into account — re-plan, offer alternatives, or escalate to a human — rather than repeating itself?

**In this project we aim to...**
- Build a guided troubleshooting prototype for a familiar physical asset (e.g. a household appliance), driven by its real user and service manuals.
- Implement and evaluate the draft → expert verification → fixed instruction set pipeline, measuring repeatability and guidance quality across repeated runs of the same fault.
- Design the boundary and escalation model (safe vs. restricted interventions, when to hand over to a professional) and the mid-procedure feedback loop.