EthosPowerAIReference design · Intelligence
Plant Training Portal on Approved SOPs
- DeepTutor
- Ollama
- Docker
- SharePoint
- Power Automate
- MinerU
- PageIndex
The problem
A pilot plant moving towards continuous production has new starters and new shift patterns to train, one IT person, and process know-how it cannot send to a public chatbot. Operators learn from whichever SOP revision is to hand, and nothing ties a training answer to the approved document.
What I would build
DeepTutor runs in Docker on a server on site with local models through Ollama, web search and every outbound feature switched off, and a packet capture at commissioning proving nothing leaves. A scheduled sync copies only the revisions the document controller approved in SharePoint into its knowledge base, removing superseded ones, and supervisors' questions must pass kb eval before anyone learns from it. Each role gets a mastery path, quizzes and cited answers; when the tutor's gate passes, a supervisor observes the task on the real equipment and records the sign-off in a SharePoint list. Questions the tutor could not ground go to the coordinator as document-control actions.
The logic flow
Click any node for what it does, what it hands on, and where a person is required.
What it returns
This is a reference design on DeepTutor, not a system I run, so there is no measured time saving to quote, and the authors' own benchmark gains on university curricula with a simulated student do not transfer to plant training. The run shown is illustrative. The numbers I would watch are kb eval scores against the supervisors' questions after each upgrade, and how many gaps per month the tutor sends back to document control.
Why it fits
An operator on a graphite anode line can now photograph a work instruction, paste it into a chatbot on her phone and get a plain-language explanation before her supervisor has finished the morning handover. Training used to be the thing a plant bought from outside, or carried in the heads of its longest-serving people, and now every new starter carries a tutor in her pocket. A plant that answers by letting that happen has sent its patented process steps to a service it does not control, and taught its operators from whichever revision of the SOP happened to be in the photograph.
So I would bring the tutor inside the fence. DeepTutor runs in Docker on a server on site, with local models through Ollama and web search and every outbound feature switched off. The project does not document what it sends by default, so a packet capture during commissioning proves that nothing leaves the host. Only the revisions the document controller has approved in SharePoint reach the knowledge base, and a superseded SOP leaves it the same day. Each role gets a mastery path built from those documents, every answer cites the page of the SOP it came from, and the supervisors write a short set of questions that the retrieval must answer correctly before a single operator learns anything from it. The tutor's mastery gate decides when an operator is ready to be watched; it does not decide that she is competent. A supervisor signs that off at the equipment, in the SharePoint training register the plant already keeps, and every question the tutor could not ground goes to the coordinator as a gap in the documents.
One five-year strategy it could serve, offered as a sketch rather than a plan: treat the gap list as the product and the tutor as the instrument that finds it. A plant moving from batches towards continuous production will rewrite most of its procedures more than once, and a record of which questions operators asked that no approved document could answer tells the quality lead where the next rewrite should start, long before an audit or an incident on the line does, and the operators add to that record every time they ask the tutor something it cannot answer. Any chatbot can explain an SOP. Only the plant can make the SOP right.
See it run
Want one of these?
Every part of it is open source and self-hostable. What I bring is the build, the checks around the project's gaps, and the judgement about where your gates should sit.