CAMPUX learn / ai
Free. Microsoft Foundry and Azure.
Ten projects, two fictional clients.

Everyone has a chatbot demo. Nobody got hired for one.

Projects you build, break on purpose, and can explain when the tool call fails and someone asks who paid for the tokens.

projects
10 live. Each built and torn down in a real subscription before it ships.
platform
Microsoft Foundry and Azure, current names, preview status marked where it applies.1
client
Campux Retail, a fictional appliance retailer. Every project solves one of its problems.
exams
Overlaps AI-103 and AI-200. Not exam prep.
updated
September 2026

The job you are actually applying for

A composite of AI engineering postings from this summer. Two lines are circled. Neither is "familiar with ChatGPT".

composite posting, 2026

AI Platform Engineer (Azure) · Hybrid · Senior

  • Build and operate LLM-backed services on Microsoft Foundry or Azure OpenAI
  • Design agent workflows that call tools, keep state, and recover from failures
  • Secure endpoints with Entra ID, managed identity, and RBAC
  • Instrument agents with tracing and evaluation so we can see why an answer was wrong
  • Explain cost, latency, and model choice trade-offs to non-engineers

the circled ones get asked about in the second interview

What an agent actually is

A loop. The model reads the conversation, answers or calls a tool, the result goes back in, repeat.

The agent loop as deployed: a customer signs in to a Streamlit app on Azure Container Apps, which calls an agent in a Microsoft Foundry project; the agent uses a model, tools that reach your systems, per-user memory, and exports traces to Application Insights; Microsoft Entra ID is the identity plane. the tool result goes back into the conversation. That is the loop. CLIENT MICROSOFT FOUNDRY PROJECT YOUR SYSTEMS AND MONITORING Customerweb browser Container AppsStreamlit UIEntra sign-in Agent · Foundry Agent Serviceconversation in, response out · the loop lives here Modelanswer, or call a tool Memoryper user, across turns Toolsfile search · OpenAPI · Azure Functions · MCPthe result goes back into the conversation Your systemsorders · warranty · catalogreached only by the agent Monitoringtraces · evals · costApplication Insights Microsoft Entra IDevery hop is an identity, never a key 1 2 3 4 5 6 7
livepress play to watch one question go round the loop
Figure 1 The loop every project is built on, drawn the way it is deployed. Green tiles are the Foundry objects you create by name in Project 01.
  1. Sign in. The customer opens the Streamlit app on Azure Container Apps and signs in with Microsoft Entra ID.
  2. Call the agent. The app calls the agent as the customer, never with a shared key.
  3. Reason. The model reads the conversation and decides: answer, or call a tool.
  4. Use a tool. The agent calls a tool and the result goes back into the conversation. That is the loop.
  5. Reach your systems. Tools touch orders, warranty and catalog data; the customer never does directly.
  6. Remember. The turn is written to per-user memory so the next one has context.
  7. Trace it. Every hop is exported to Application Insights: traces, evaluations, cost.

The projects

Each is one page: the problem, the architecture, a numbered build where every step ends with a check, where it breaks, the trade-offs, and what to say in the interview. Start at 00 if the Azure portal is new to you.

  1. 00 Foundry in an afternoon Your first agent, portal only. A Foundry project, a deployed model, and an agent that answers from three documents and admits when it cannot. Ninety minutes, nothing installed, everything deleted at the end. live
  2. 01 Warranty Desk A deployed support agent that looks up orders, checks warranty, answers from policy documents, remembers the customer, and shows you its own traces. live
  3. 02 Ground it Retrieval with Foundry IQ and Azure AI Search. The same agent answering from thousands of documents without making things up, with a measured hit rate. live
  4. 03 Evaluate and red-team An evaluation set, a scored run, a prompt-injection attempt that fails, and a dashboard you can show a manager. live
  5. 04 Multi-agent workflow Microsoft Agent Framework. A triage agent handing off to specialists, with a human approval step that actually blocks. live
  6. 05 Ship it like infrastructure Bicep, pipelines, private networking. The whole stack as code, deployed from GitHub Actions into a private-endpoint-only environment. live
  7. 06 Voice desk A phone-style agent with Voice Live. The support agent you can talk to, with latency you had to fight for. live
  8. 07 Care Gap Report A HIPAA-aware quality agent for Campux Health, a regulated clinic network. Redacts identifiers before any model call, cites every decision to a source value, gates re-identification behind a human, and logs everything. live
  9. 08 Prior Authorization A payer-policy agent for Campux Health. Matches a clinical order to a payer's medical-necessity criteria with cited evidence, refuses to draft unsupportable requests, and routes every submission through a clinician. live
  10. 09 Medication Reconciliation Cross-check one patient's medication list across two record sources for Campux Health, flag dose, duplicate and interaction discrepancies with both sources cited, and hand a pharmacist the report to sign. live
Start at 00 Know Azure already? Start at 01 Platform engineer? Start at 05

What you need

An Azure subscription you control
Free trial or pay-as-you-go. Every project ends with a teardown step. Do it.
Python and a terminal
Python 3.11 or later, the Azure CLI, and the Azure Developer CLI (azd). Project 00 needs none of this, only a browser.
Four words you can explain, from Project 01 on
Resource group, Entra ID, managed identity, role assignment. Project 00 introduces the first two. The CAMPUX bootcamp covers all four, but it is not required here.

Notes

  1. Microsoft folded Azure AI Foundry into "Microsoft Foundry" in 2026. The projects use the current names and say so where a feature is still in preview.