Research

Applying AI where a factory can actually trust it.

We treat AI as a tool for surfacing and ranking evidence a human can verify — never as a black box that hands down a verdict. Every area below exists because a real investigation needed it, not because it was the interesting thing to build.

What we're working on

Evidence-grounded hypothesis generationRanking candidate root causes from production, quality and maintenance history, with every claim traceable to the record it came from — not an opaque confidence score.

Operational signal correlationLinking a new quality event, downtime record or production anomaly back to an open investigation by matching product, line, asset, reference and severity — so related signals surface automatically instead of staying siloed by module.

Statistical process controlControl-chart limits, out-of-control detection and automatic non-conformance creation from live subgroup measurements — classical SPC, applied where it belongs, not reinvented.

Recurring-failure and recurring-loss detectionSurfacing the failure codes and loss patterns that keep coming back — on the same line, the same product, the same asset — before they're big enough to force attention on their own.

Grounded retrieval for operator questionsA knowledge assistant that answers factory questions by retrieving and citing your own SOPs, manuals and investigation history, with live data lookups where a static document can't answer — not a model guessing from general training.

Our approach: provenance first

A finding is only as useful as a quality manager's ability to check it. Every hypothesis, correlation and recommendation Qmach-AI surfaces carries a citation back to the underlying quality event, production record, maintenance ticket or document it was derived from. We would rather ship a smaller, verifiable claim than a bigger, unverifiable one.

Local-first, tenant-isolated inference

Where the platform uses a language model — for hypothesis phrasing, document verification, or the Nut AI assistant — it is built to run against a self-hosted model server, keeping factory data inside the tenant's own boundary rather than sent to a third-party API by default. Every AI-assisted feature also degrades gracefully: if the model is unreachable, the deterministic, evidence-only output still ships — a summary is never withheld just because an LLM call failed.