Case study / Manufacturing
An AI assistant that answers engineers' questions from the manuals and the breakdown history, shows exactly where each answer came from, and says so when it does not know.
01 / The industry
Factories run on machines that stay in service for decades. The manuals run to hundreds of pages of PDF, and the record of what went wrong, and what fixed it, sits in breakdown logs and spreadsheets that nobody reads.
The real knowledge is with a few senior engineers: which setting stops a line jamming, which part always fails first, what a particular noise means. When one of them retires, moves on, or is simply busy on another line, newer technicians wait, guess, or ring the manufacturer.
02 / The problem
A general-purpose chatbot is the wrong tool here. It answers confidently when it does not know, and on a machine a wrong answer can damage equipment or hurt someone.
The brief was an assistant that:
03 / What we built
Engineers ask questions in their own words, on a desktop or a tablet, usually about a particular machine. A printed QR label on the machine opens a conversation about it. If a question could mean more than one machine, the assistant asks which.
A second view is for the people responsible for the knowledge itself. Repeat questions show where training is needed. Questions the assistant could not answer show the gaps, machine by machine. Documents that are voted down, queried or never cited are flagged for review.
Every document has an owner and a review date, following the document control approach in ISO 9001, with an export to show an auditor. A withdrawn document stops being used for answers.
04 / How it works
Each question is searched two ways at once: by meaning, and by exact words, so part numbers and fault codes are found as reliably as descriptions. The best matches are re-ranked, and a confidence check runs before any answer is written.
It runs on AWS in London, with the infrastructure defined in Terraform. Documents are read inside the platform rather than sent to a third-party document service, each organisation's data is kept separate at the database level, and actions are recorded in an audit log.
A benchmark of test questions, including ones it must refuse, measures answer quality as the system changes, alongside more than 3,000 automated tests.
05 / Where it is now
The assistant has been piloted on real material: equipment manufacturers' manuals and years of a maintenance team's breakdown records. In the pilot evaluation:
Next comes the production environment, reading scanned documents, and a mobile app for technicians on the shop floor.
06 / More work
Vehicle leasing
A vehicle leasing provider ran salary sacrifice schemes on a CRM, with manual steps at every stage. We built the pricing engine, portal, partner integrations and automation that take a deal from first quote to monthly mileage.
Commercial fit-out
A commercial fit-out contractor's project and finance data sat in a decade-old system with fixed reports. We built live dashboards into it for forecasting, work in progress and daily changes, plus wallboards in its offices.
Construction
A building contractor's site teams often had no signal on the job, so updates went by WhatsApp and email. We built an iOS and Android app on its existing back office that works offline and syncs when it can.
07 / Contact
Tell us where the know-how in your business lives today, and what happens when the people who hold it are not there.