Industrial AI for engineering work, with its limits in code.
Creweng's AI division builds agents, knowledge graphs, harnesses and document pipelines for utilities, water and wastewater, heat, plant and data centres, and for the consultancies and contractors that serve them. It also builds and runs the agent fleet and document pipeline behind Creweng's own briefs, dossiers and design packs: Creweng Water & Heat works through that pipeline, and Industrial AI Assurance tests and evidences AI connected to plant, including systems the division did not build. Fixed price, quoted after a short scoping call.
How the work is done
- Your provider, your keys, your accounts
- Every answer cites its source
- Read-only near plant, no control write-back
- Tests gate every deploy
- Drafts reviewed by an engineer before release
- Controls mapped, never a status claimed
The problem
Engineering runs on documents: drawings, permits, design packs, operating manuals and the rules behind them. A chat window over a folder of files cannot say which revision it read, keeps its limits in a prompt, and nobody has checked what it can reach. The division builds the parts that make AI usable on engineering work: a graph of the documents and the duties, agents whose limits are code and tests and that stay read-only near plant, and a pipeline that drafts a document set for an engineer to review. It runs the same method on Creweng's own work first.
The answer has no source
An assistant that cannot point to the drawing, the clause or the revision it read gives an answer nobody can check or sign. Every answer has to cite the document and the version it came from.
The limit is a prompt
The rule that keeps an agent away from the controls lives in a system prompt, so it holds until the first unexpected input. A limit has to be code, checked before the model acts and pinned by a test that gates every deploy.
Every tool keeps its own copy
The design team's assistant, the operator's assistant and the compliance tracker each hold a private copy of the same documents and duties, and they drift apart. One shared graph gives every agent the same context and says which version it is.
Each document set starts from blank
Briefs, dossiers and design-pack sections are rewritten on every job and cross-checked by eye. A pipeline drafts the set from a requirements sheet, keeps the register and the revision log, and leaves the judgement to the engineer who reviews it.
What you walk away with
The products, in plain words.
- Engineering agents: assistants and drafting agents over your own manuals, drawings, permits and records. Each answer cites the document and revision it used, access is checked inside the retrieval query, and near plant they are advisory and read-only, with no control write-back.
- Knowledge graphs: documents, duties, assets and records modelled as one graph your agents and your people share, with retrieval measured on your own questions and re-measured after every change, and MCP access for your own tools.
- Harness engineering: tool schemas, permissions, hooks, context and memory, budgets, approval points and logs, in your account, with tests that gate every deploy.
- Document pipelines: from a requirements sheet or a site profile to a drafted document set (briefs, dossiers, design-pack sections) with a document register and a revision log, delivered for an engineer to review and adopt.
- Compliance engineering: controls engineered against the rules that apply to the system (UK GDPR, the GDPR, the EU AI Act) and mapped in an evidence register. We describe the control; we never claim a status.
- Two fixed-price engagements to start with: a two-week Harness Audit of an existing system, and a Minimum Viable Harness build for a new one.
What we do
Engineering agents
Assistants and drafting agents over your own documents and data that cite what they read, and stay advisory and read-only near plant.
Knowledge graphs
Documents, duties, assets and records as one graph every agent shares, with retrieval measured on your own questions.
Harness engineering
The scaffolding that governs an agent: tool schemas, permissions, hooks, context and memory, budgets and approval points.
Document pipelines
Requirements in, a reviewable document set out: the drafts, the document register and the revision log, with an engineer's review as the last step.
Compliance engineering
Controls engineered against the rules that apply and mapped in an evidence register: the control is described, never a status claimed.
Creweng's own fleet
The agents and the pipeline behind Creweng's briefs, dossiers and design packs, built and run by the division on the method it sells.
How we engage
Three steps, one fixed price.
- 01
Scoping call
The work, the documents and data, any plant interfaces and the accounts it will run in. A fixed price in writing follows.
- 02
Build or audit
The agents, the graph, the harness and the pipeline, or an audit of the ones you have, with every deploy gated by tests.
- 03
Run or hand over
Keep the division on retainer to run it, or take it in-house: everything is code and tests in your repository.
Where it matters most
Who buys it, and when.
A utility or an operator with an archive that is hard to search
An assistant over the manuals, drawings, procedures and permits you already hold, answering with the document and revision, with access checked inside the query and no write-back to plant.
A consultancy or contractor that writes the same documents on every job
A pipeline that drafts the recurring set from your own templates and requirements, keeps the register and the revision log, and leaves the engineering judgement with your reviewers.
A data-centre developer or a heat network tracking duties across its sites
The obligation graph behind Creweng Water & Heat, built for your own portfolio, with MCP access so your own agents read the same duties and sources.
A plant-AI vendor before a customer's pilot
The harness, the evaluation and a written statement of what the system may not do, in place before it reaches a customer's plant. Industrial AI Assurance can then test it, with an audit-independence disclosure because the division built it.
Beyond engineering
The division's other AI work on one page: harness engineering, data graphing, compliance engineering, adversarial testing, integrations and governed delivery, and the agents and tools it has built.
Questions
Tell us what you are deciding.
A short scoping call, then a fixed price in writing.