AI-Augmented Python and Azure for Manufacturing
Plant data to a predictive model to an advisor agent, on Azure
About this course
A manufacturing-specific AI engineering course. It follows one thread the whole way: machine telemetry and plant systems land in a data lake, a failure-prediction model is trained and served from it, retrieval over manuals and SOPs turns a prediction into a recommendation, and an agent hands a planner a cited, evidence-backed next action.
Nine modules across three parts — plant data and prediction, generative AI and agents, then integration and deployment — ending in an integrated case study in which you build the whole thread yourself.
The build uses Azure IoT Hub, Data Lake Storage, Azure Machine Learning, Azure AI Foundry and Azure AI Search, with Python throughout. Lessons focus on the architectural decisions so they stay useful as cloud products evolve.
What you’ll be able to do
- Read a manufacturing data landscape and say what PLM, MES, ERP, IoT and unstructured knowledge each can and cannot tell you
- Land telemetry and batch system extracts into a medallion data lake and join them correctly in Python
- Frame a failure-prediction problem with a horizon a maintenance planner can actually act on, and label it honestly
- Train, explain, register and serve a failure-prediction model, and evaluate it against the cost of a false alarm versus a missed failure
- Ground a generative model on manuals, SOPs and alarm-code tables so answers carry citations
- Build an advisor agent with guardrails, human approval and no direct writes to systems of record
- Deploy the whole prototype, monitor it, and present it as a one-page solution brief with a business case
Before you start
What this course assumes you can already do.
- Working Python: pandas, writing a module, running a script from the command line.
- Basic machine learning literacy — what a classifier is, what precision and recall mean.
- An Azure subscription you can create resources in, and familiarity with the portal.
- Helpful but not required: exposure to plant systems (MES, ERP, PLM) or industrial telemetry.
The curriculum
Open any module to see its objective, every teaching topic, the practice brief and the checkpoint.
Recaps and reference material are included in the LMS.
PART 1Manufacturing data and predictive AI with Python
01Manufacturing and PLM Context for AI
Be able to sit in front of a manufacturing customer and say where AI creates value across their lifecycle, what their systems can actually tell you, and how to frame one use case eight ways before promising anything.
- Where AI creates value across the lifecycleConcept
- The manufacturing data landscapeConcept
- Framing a use case from the customer perspectiveConcept
- Use case briefing: predictive maintenance to prescriptive actionConcept
Map one plant you know, then frame one use case eight ways
Take a real plant, line or product you have visibility into. Fill in the manufacturing data landscape for it, then put one candidate use case through the eight-part frame from this module.
Knowledge checkpoint with explained answers
02Manufacturing Data on Azure - IoT Hub and Data Lake
Get real telemetry from a device into a medallion data lake, join it with plant system extracts in Python, and handle the data problems every plant actually has.
- Reference architecture: edge to lake to analyticsConcept
- Hands-on: telemetry to lakeHands-on
- Tech stack and plant data realitiesHands-on
- Production analytics from the same dataHands-on
Join telemetry to plant systems and survive the mess
Stream simulated machine telemetry into a lake, land a batch extract beside it, and join them. Then inject the failures every plant has — drift, gaps, duplicates, late messages, asset IDs that do not match — and make the join hold.
Knowledge checkpoint with explained answers
03Predictive Maintenance Models with Python
Turn "the machine will fail" into a labelled, evaluated, explainable and served prediction with a horizon the maintenance planner can act on — and then into a number the business recognises.
- Framing the predictionConcept
- Hands-on: failure-prediction modelHands-on
- Tech stack for the prediction pathHands-on
- From prediction to business metricConcept
Pick a horizon a maintenance planner can act on
Train a failure-within-N-days classifier and then argue N. Try at least two horizons and show what changes for the planner and for the model.
Knowledge checkpoint with explained answers
PART 2Generative AI, RAG and agents over manufacturing knowledge
04Azure AI Foundry Foundations
Set up a Foundry project with model access governed properly, prompt for reliable structured behaviour, and turn a prediction into a work-order draft a planner would read.
- Foundry set-up: hubs, projects, model accessHands-on
- Prompting for reliable behaviourConcept
- Hands-on: prediction to work-order draftHands-on
- Tech stack for the generative pathHands-on
Make the work-order draft trustworthy
Prompt a deployed model to turn a prediction and its top features into a structured maintenance work-order summary. Then attack it: missing features, contradictory signals, a prediction with low confidence.
Knowledge checkpoint with explained answers
05RAG over Enterprise Manufacturing Knowledge
Index the messy real knowledge of a plant — manuals with tables, alarm-code lists, technician notes — and answer an operator question with the governing document and section cited.
- RAG essentials for manufacturing documentsConcept
- Hands-on: equipment knowledge assistantHands-on
- Tech stack and measuring retrieval qualityHands-on
Answer an alarm-code question with the SOP cited
Index a document set that mixes formats — a manual with tables, an alarm-code list, SOPs, free-text maintenance notes — then build a gold question set and measure retrieval before you judge the answers.
Knowledge checkpoint with explained answers
06Agentic AI - From Prediction to Recommendation
Build the advisor agent: tools that reach the model endpoint, the knowledge index and plant systems, guardrails that keep it out of systems of record, and evidence attached to every recommendation.
- Single agent: role, instructions, tools, memoryConcept
- Agent tools and multi-agent orchestrationConcept
- Guardrails and human-in-the-loop approvalConcept
- Hands-on: maintenance advisor agentHands-on
- Tech stack for the agentic pathHands-on
Give the advisor agent guardrails it cannot talk its way past
Build the advisor agent with tools onto the prediction endpoint, the knowledge index and mocked plant APIs. Then try to make it do something it should not.
Knowledge checkpoint with explained answers
PART 3Integration, deployment and the end-to-end build
07Integration and Deployment in the Customer Ecosystem
Connect the prototype to the customer ecosystem without becoming a risk to it: integration patterns, where each component lives, monitoring and drift, and the security controls that make it allowed.
- Integrating with manufacturing systemsConcept
- Hosting, scaling and maintainingConcept
- Hands-on: deploy the prototypeHands-on
- Tech stack for hosting and operationsHands-on
Deploy it as though someone else will operate it
Containerise the advisor application and deploy it with managed identity, connected to the model endpoint, the agent and the search index. Then hand it over on paper.
Knowledge checkpoint with explained answers
08Case Study: Predictive Maintenance to Prescriptive Action
Build the whole thread in one block — data to prediction, prediction to recommendation, deployment and integration — then score it against a gold question set and present it.
- Case study: the integrated build blockConcept
- Solution design on one pageHands-on
- Build 1 — data to predictionHands-on
- Build 2 — prediction to recommendationHands-on
- Build 3 — deployment and integrationHands-on
- Evaluation and showcaseHands-on
Run the integrated build against the clock
Work the case study in its planned stages: solution design, then the three builds, then evaluation and a showcase. The point of the time pressure is to force the design decisions to be made early and written down.
Knowledge checkpoint with explained answers
09Taking It to the Customer
Convert a working prototype into a proposal: the one-page brief, a business case built on baseline metrics, and the same frame applied to the next use case.
- From prototype to proposalConcept
- Building the business caseConcept
- Applying the frame to other lifecycle use casesConcept
Write the proposal and the second use case
Turn the prototype into a proposal a customer could act on, then apply the same eight-part frame to a different lifecycle use case to prove the frame travels.
Knowledge checkpoint with explained answers
Predictive maintenance to prescriptive action
The course case study, run as the capstone: one integrated build across the whole stack, from plant telemetry to a cited, evidence-backed recommendation on a planner's screen, deployed and presented as a proposal. Built on synthetic or anonymised plant data only.
Solution design (one page) — Commit the design to paper before building anything.
Build 1 — data to prediction — From lake to a registered, explained model behind an endpoint.
Build 2 — prediction to recommendation — Retrieval over plant knowledge plus an advisor agent that returns evidence, not opinion.
Build 3 — deployment and integration — Deployed, monitored, and connected through labelled stubs.
Evaluation and showcase — Score it against the gold set, then present it as a proposal.
Next course
Tell us about yourself or your team and we’ll let you know how and when you can take this course.