Which problem and metric does this address, and can you show a named reference in a similar process?
DatumKeep closes the knowledge gap plant engineers face when they resolve incidents: it finds how similar problems were solved before, even when the engineer has no prior experience with them. It proposes hypotheses and actions within a guided 8D, with 5 Whys and Ishikawa at the root-cause step. DatumKeep measures time to close an investigation, incidents per equipment and failure mode, and process capability (Cpk); business metrics such as cost of poor quality or OEE we calculate with you from your data. We have no named reference to share yet, so we propose a time-boxed pilot with a decision rule agreed in advance: scale, adjust or stop.
Where is our data stored and processed, and is it ever used to train models shared with other customers?
We do not train models on customer data. Evaluating models with your data happens only at your request. Customer pilots run on EU-hosted models (Amazon Bedrock); our demo and trials use the Anthropic and OpenAI APIs directly. Application and database are hosted in the EU; EU file storage is being migrated.
How is each output traced back to a source, and where does a human review it?
Sources. Hypotheses and actions show the past case they come from, when there is one, and similar cases link to the original investigation.
Numbers. Statistics are calculated by statistical engines; reports cannot claim a capability figure that was never calculated.
Human review. The AI proposes and pre-scores hypotheses. The engineer confirms or rejects each one, can add their own, and can rewrite the root cause. The engineer writes their own actions or accepts the proposed ones; AI-proposed actions are saved only once the engineer accepts them. The engineer closes the investigation and signs the report.
What happens when the model is wrong or unavailable? Does the system degrade honestly?
DatumKeep keeps working when the AI is unavailable: investigations, statistics and records stay usable, with fallbacks where a model is involved, and a failing AI provider is isolated so it does not block the rest. Every statistic comes from a statistical engine. In the SPC assistant, AI text containing a number no engine returned is discarded, and reports cannot claim results the record does not support. A contradiction check flags conflicting statements, and with little data DatumKeep says what is not yet proven.
How do you measure return, and can the contract be tied to an agreed result?
DatumKeep measures time to close, recurrence per equipment and failure mode, process capability (Cpk), and each action's effect through statistical before/after comparisons, stating when the data is still insufficient. The pilot is paid and has success criteria agreed in advance. Met: it extends to a licence with a negotiated scope. Positive with conditions: we adapt DatumKeep, then extend. Not met: we analyse the cause with you, fix it and agree whether to rerun.