TWO WORLDS, NO BRIDGE
Sensor and historian data never joins cleanly to ERP and quality records, so nobody can say which order the scrap came from.
USE CASE / MANUFACTURING
The machines know what happened. The ERP does not.
Shop-floor data and business data grow up in different worlds. One is a stream of sensor readings every few seconds. The other is a set of orders and part numbers. Joining them is where the value is, and it is also the part almost nobody has done properly.
What usually goes wrong
OEE LIVES IN A SPREADSHEET
The single most important number in the plant is calculated by hand, differently in each plant, so the plants cannot be compared.
QUALITY PROBLEMS ARRIVE LATE
A defect pattern is spotted weeks after it started, once the batch has already shipped.
MAINTENANCE IS REACTIVE
The line stops, then people start looking. The knowledge of what usually causes it sits in one engineer's head.
MANUALS ARE UNSEARCHABLE
Thousands of pages of machine manuals and standard procedures exist as PDFs that nobody can search when the line is down.
What we build
- Streaming ingestion from historians, PLCs and IoT gateways into a lakehouse, without dropping readings when the network hiccups.
- A joined model that links a sensor reading to a work order, a batch, a part number and a quality result.
- One agreed OEE calculation, defined once and applied identically across every plant, so plant comparison is finally honest.
- Data quality tests on sensor feeds, so a stuck sensor reporting the same value all day is flagged instead of quietly averaged in.
- Time-aligned data, so a reading taken at 14:03 on one machine lines up correctly with a reading from another machine in a different time zone.
- Dashboards that a shift supervisor can read in ten seconds on a screen on the factory floor.
What AI looks like here
Line stoppage explainer
- What it does
- When a line stops, the assistant writes a first-draft explanation: what the sensors were doing before the stop, whether this has happened before, and what fixed it last time.
- How it works
- The platform assembles the sensor history for the minutes before the stop and finds similar past incidents using embeddings. That evidence, plus the relevant section of the maintenance manual, goes to a language model, which writes a short summary with links to what it used.
- What it needs
- Reliable streaming sensor data, a history of past incidents with what resolved them, digitised manuals, and a vector store. The output is always a draft for a human engineer to confirm.
Maintenance manual assistant
- What it does
- A technician on the floor asks a question in normal language and gets the exact procedure from the right manual, on a phone or tablet, in their own language.
- How it works
- Manuals and procedures are converted to text, split into sections, and stored as embeddings. The question retrieves the closest sections, and the model answers using only those sections and shows the page it came from. Answers can be translated into the languages spoken in the plant.
- What it needs
- Digitised manuals — scanned PDFs need OCR first — a vector store, and a strict rule that the assistant refuses to answer when it cannot find a relevant section.
Weekly defect pattern summary
- What it does
- Groups the week's defect records by what they actually have in common, and writes a short summary of the top recurring themes for the quality meeting.
- How it works
- Free-text defect notes are turned into embeddings and clustered, so notes describing the same real problem in different words end up together. A model then names each cluster and writes one paragraph about it, with the record count and the machines involved.
- What it needs
- Defect records with usable free-text notes, a link from defect to machine and batch, and a person who reviews the summary before it is used in a decision.
The tools we use here
- Databricks with Delta Lake
- Spark Structured Streaming
- PySpark
- Azure IoT Hub, Event Hubs or Amazon Kinesis
- dbt
- Apache Airflow
- Terraform
- Databricks Mosaic AI or Amazon Bedrock
- Unity Catalog
- Power BI
What changes for you
- One OEE number that every plant calculates the same way.
- A stoppage gets a first explanation in minutes instead of a meeting the next day.
- Technicians find the right procedure without walking back to the office.
- Quality patterns surface in the same week they start, not the month after.
What we will not do
- We do not connect anything to a control system that can change machine behaviour. Read only.
- We do not let a model approve a batch, sign off quality, or authorise a repair.
- We do not deploy an assistant that guesses when the manual does not cover the question.
- We do not replace an engineer's judgement. We give them a faster first draft.