AI Instructor Live Labs Included Exam Prep

AI-3016: Develop generative AI apps by using Microsoft Foundry (APL-3016 prep)

Rebuild the Halcyon Insurance claims copilot end-to-end on Microsoft Foundry and prep for the APL-3016 Applied Skills assessment: account + project + connections, model deployment, Microsoft Agent Framework, RAG, responsible AI, evaluation, and a capstone.

Intermediate
16h 3m
11 Lessons
AI-3016
AI-3016: Develop generative AI apps with Microsoft Foundry (APL-3016) Badge

View badge details

Exam Preparation

Exam Preparation Included

Practice with real exam-style questions for the AI-3016 certification. AI-powered feedback helps you understand every answer.

Practice Questions Track Progress AI Feedback

About This Course

Ship a production-quality generative AI copilot on Microsoft Foundry and prep for the APL-3016 Applied Skills assessment. Over 11 lessons you'll rebuild the fictional Halcyon Insurance claims-adjuster copilot end-to-end Foundry account + project + connected resources, model deployment, Microsoft Agent Framework orchestration, RAG grounding against a policy-document library, custom content-safety filters, quality + risk-safety evaluation, and a full end-to-end capstone. Every artifact is grounded against the current Microsoft Foundry product (AIServices topology, Agent Framework replacing Prompt Flow, current model versions).

Course Curriculum

11 Lessons
01
AI Lesson
AI Lesson

Microsoft Foundry: what it is and what you'll build

1h 0m

Iris introduces the Foundry topology: a Microsoft.CognitiveServices/accounts resource with kind: AIServices and allowProjectManagement: true, Foundry projects as sub-resources, and typed connections to Storage and AI Search. She also maps the renamed vocabulary from AI Studio and AI hubs to the New Foundry portal.

Students leave able to describe the exact Foundry stack Lab 2509 provisions for Halcyon Assist.

02
Lab Exercise
Lab Exercise

Set up your Foundry account, project, connections, and models

2h 3m 7 Exercises

Students provision the Halcyon Assist Foundry stack by hand. Using the Azure CLI, they run az cognitiveservices account create with kind AIServices, then az cognitiveservices account project create for halcyon-project. They deploy gpt-5.1 and text-embedding-3-large, and add AI Search plus Storage connections.

After this lab, students can stand up a Foundry project from scratch that later labs build on.

Verify your Azure sign-in and set your lab variables ~8 min
Create the Foundry account ~10 min
Create the Foundry project ~10 min
Connect Azure Storage to the Foundry project ~12 min
Connect Azure AI Search to the Foundry project ~12 min
Deploy the chat and companion models ~12 min
Deploy embeddings and smoke-test from Python ~14 min
03
AI Lesson
AI Lesson

Copilot patterns with Microsoft Agent Framework

1h 0m

Iris teaches the five-layer copilot mental model — Grounding, Model, Orchestration, Tools, Session — and maps each layer to Microsoft Agent Framework primitives: FoundryChatClient, Agent, AgentThread, function tools, and streaming via agent.run_stream(...). She walks through how Halcyon Assist will be built as an Agent Framework agent.

Students leave ready to wire the code in Lab 2511 and know why each layer exists.

04
Lab Exercise
Lab Exercise

Build the Halcyon Assist copilot

2h 5m 6 Exercises

This lab pre-deploys resources and may take 15 or more minutes to start.

Build the first working version of the Halcyon Assist claims-adjuster copilot against a pre-deployed Foundry stack (account, halcyon-project, gpt-5.1, gpt-5-mini, text-embedding-3-large). Students author the system prompt, wire AIProjectClient, rewrite in Microsoft Agent Framework, add a lookup_policy_summary function tool, and enable streaming with AgentThread.

Verify the pre-deployed Halcyon Foundry stack ~10 min
Author the Halcyon Assist system prompt ~10 min
First working conversation with AIProjectClient ~15 min
Rewrite with Microsoft Agent Framework ~15 min
Add a function tool: lookup_policy_summary ~15 min
Streaming and multi-turn threads ~15 min
05
AI Lesson
AI Lesson

Retrieval-Augmented Generation: grounding LLMs against your data

1h 0m

Iris explains why LLMs hallucinate and how Retrieval-Augmented Generation fixes it. She covers vector, keyword, and hybrid retrieval; Azure AI Search integrated vectorization; the semantic ranker; embedding models like text-embedding-3-large; chunking strategies; and citation patterns that let a copilot cite its sources.

Students leave ready to ground Halcyon Assist against the policy library in Lab 2513.

06
Lab Exercise
Lab Exercise

Add RAG to Halcyon Assist

2h 15m 6 Exercises

This lab pre-deploys resources and may take 15 or more minutes to start.

Ground Halcyon Assist against Halcyon Insurance's policy library. Students upload homeowners, auto, and umbrella PDFs to the pre-created halcyon-policies blob container, build an integrated-vectorization index through the Foundry portal Import wizard using text-embedding-3-large, wire get_azure_ai_search_tool onto the Agent Framework agent, and compare grounded versus ungrounded answers.

Verify the pre-deployed Halcyon RAG stack Confirm the ARM-deployed Foundry account, project, model deployments, Storage account, and AI Search service are all in place and reachable, and that the starter project loads cleanly in the VS Code container. ~10 min
Upload the Halcyon policy library to Blob storage Batch-upload the three Halcyon policy documents from the starter project's data/policies/ folder into the pre-created halcyon-policies Blob container so the AI Search indexer has a corpus to index. ~10 min
Create an AI Search index with integrated vectorization (Foundry portal) Use the Foundry portal Import wizard to create an integrated-vectorization index over the halcyon-policies blob container, using text-embedding-3-large from the Foundry chat account as the embedding model. ~20 min
Verify the index — query from the Foundry Search playground Run one hybrid-semantic query against halcyon-policies-index in the Foundry portal Search playground and confirm the wizard-generated index returns the expected policy chunk with a semantic re-ranker score. ~10 min
Wire the ai_search tool onto the Halcyon Assist agent Edit src/rag_client.py at the ai_search-tool-wiring marker: resolve the Foundry Search connection ID, attach FoundryChatClient.get_azure_ai_search_tool to the agent, and run one end-to-end grounded query from the lab-bash terminal. ~25 min
Grounded vs ungrounded — compare three test queries Run three Halcyon customer-service questions against both the ungrounded lesson-4 agent and your newly grounded lesson-6 agent, then discuss which questions benefit from grounding and why. ~15 min
07
AI Lesson
AI Lesson

Responsible AI in Foundry: filters, blocklists, and safety policies

45m

Iris teaches Foundry's Responsible AI controls: input and output content filters for Hate, Sexual, Violence, and Self-harm; blocklists with exact and regex matching; protected-material detection for text and code; and the raiPolicies ARM resource type. She contrasts the default policy shipped with every deployment against custom policies with tighter thresholds.

Students leave ready to author and attach a custom filter in Lab 2515.

08
Lab Exercise
Lab Exercise

Custom content filters for Halcyon Assist

1h 0m 5 Exercises

This lab pre-deploys resources and may take 15 or more minutes to start.

Students author a custom Responsible AI policy for Halcyon Assist in the Foundry portal: an input blocklist for competitor names and a PII regex, output filters with protected-material detection for text and code, and tightened Hate/Sexual/Violence thresholds. They attach the raiPolicy to the gpt-5.1 deployment and verify the guardrails from Python end-to-end.

Verify the pre-deployed Halcyon filters stack Confirm the Foundry account, project, gpt-5.1 deployment, and the pre-provisioned example raiPolicy resource were provisioned by the lab template, and capture the endpoints into the starter's .env. ~8 min
Create a custom content filter in the Foundry portal Walk through Guardrails + controls -> Content filters -> Create in the Foundry portal to author the halcyon-assist-filter configuration with Medium severity thresholds on the four default input + output categories. ~12 min
Add an input blocklist for competitors and PII patterns Create a Halcyon input blocklist that combines exact-match competitor names (SafeguardAI, InsureSecure, ClaimsRight) with regex patterns for SSN, US phone, and email, then attach it to the halcyon-assist-filter configuration as a prompt-side blocklist. ~12 min
Add an output filter with protected-material detection Turn on Protected material for text (blocking) and Protected material for code (annotate-only) on the output side of the halcyon-assist-filter configuration and reason about how each shape shows up in the API response. ~10 min
Apply the filter to gpt-5.1 and test from Python Attach halcyon-assist-filter to the gpt-5.1 deployment (Foundry portal or az rest), then from the apl-3016-copilot starter call the model with a blocklisted prompt and a normal prompt to see both response shapes end-to-end. ~18 min
09
AI Lesson
AI Lesson

Evaluate copilot performance

45m

Iris teaches how to prove a copilot is production-ready. She covers building a golden dataset, running Foundry's quality evaluators (Relevance, Groundedness, Coherence, Fluency), layering risk and safety evaluators, reproducing runs with the azure-ai-evaluation SDK, and combining aggregate scores with manual side-by-side review.

Students leave ready to assess Halcyon Assist and pick weak categories to iterate on in Lab 2517.

10
Lab Exercise
Lab Exercise

Assess Halcyon Assist

2h 10m 6 Exercises

This lab pre-deploys resources and may take 15 or more minutes to start.

Students run a full evaluation cycle on Halcyon Assist. From a pre-seeded halcyon-eval-golden blob container, they capture a golden dataset, run Foundry portal evaluations against Relevance/Groundedness/Coherence/Fluency plus risk and safety evaluators, reproduce the run with azure-ai-evaluation, and do a v1-vs-v2 side-by-side to identify weak categories.

Verify the pre-deployed stack and capture a golden dataset Confirm the Foundry stack is up, wire the .env, then run capture_golden.py to build data/aurora-eval-golden.jsonl from 8-10 Halcyon prompts. ~15 min
Foundry portal evaluation - built-in quality evaluators Upload aurora-eval-golden.jsonl to the Foundry portal and run Relevance + Groundedness + Coherence + Fluency against your Halcyon Assist responses. ~15 min
Run risk and safety evaluators from the portal Rerun against the same golden set with Hate/Unfairness, Sexual, Violence, Self-harm, Protected material, and Indirect attack. Interpret defect-rate scores. ~15 min
Programmatic evaluation with azure-ai-evaluation Run python src/eval_runner.py to reproduce the portal evaluators programmatically. Reconcile SDK aggregate scores with the portal run. ~15 min
Manual review pass - v1 versus v2 prompt side-by-side Capture a v2 golden set from a stricter refusal prompt, then hand-review both to catch Halcyon policy violations automated evaluators miss. ~15 min
Interpret aggregate scores and identify weak categories Read the portal + SDK aggregate scores together, pick the weakest quality/safety category for Halcyon Assist, and propose one concrete next iteration. ~10 min
11
Lab Exercise
Lab Exercise

Applied Skills capstone — end-to-end Halcyon Assist

2h 0m 6 Exercises

This lab pre-deploys resources and may take 15 or more minutes to start.

End-to-end APL-3016 capstone across all seven Applied Skills tasks. Students extend Halcyon Assist with an estimate_renewal_quote function tool, extend the RAG index with renewal-terms and fee-schedule docs, author a tightened Responsible AI policy, run a full evaluation sweep, and complete the APL-3016 assessment report. Iris coaches from the sidebar but does not hand out solutions.

Verify the deployed capstone stack Inspect the pre-provisioned Foundry account, project, model deployments, AI Search, storage containers, and Application Insights that the ARM template deployed at lab start. Populate the starter .env and run verify_env.py. ~15 min
Extend Halcyon Assist with a policy-renewal-quote intent + function tool Design and register a new function tool estimate_renewal_quote(policy_id, coverage_changes) with the ChatAgent, update the system prompt so the agent routes renewal queries to the tool, and demonstrate the agent returns a formatted quote. APL-3016 tasks 2 + 3. ~25 min
Add renewal-terms and fee-schedule documents to the RAG index Extend the halcyon-policies AI Search index with renewal-terms and fee-schedule documents from the halcyon-renewal-terms blob container, re-run integrated vectorization, and verify the agent grounds renewal queries against the new corpus. APL-3016 task 4. ~20 min
Apply a tightened content filter for the renewal flow Author a custom Responsible AI policy with a stricter blocklist (competitor rates + regulatory-restricted terms), attach it to the gpt-5.1 deployment used by the renewal flow, and demonstrate a blocked-input test fires the filter. APL-3016 task 5. ~20 min
Full evaluation sweep against original + extended agent Run built-in quality evaluators (Groundedness, Relevance, Coherence, Fluency) and risk/safety evaluators (HateUnfairness, SelfHarm, Sexual, Violence) plus a small manual review against both agent versions. Produce a comparison table with metric deltas. APL-3016 task 6. ~25 min
Final APL-3016 assessment report Fill in the pass/fail assessment report covering all 7 APL-3016 tasks, linking each row to a specific capstone artifact (screenshot, log excerpt, or file path). APL-3016 task 7. ~15 min
Sign In to Get Started Start Free Trial

This course includes:

  • 24/7 AI Instructor Support
  • Live Lab Environments
  • 6 Hands-on Lessons
  • Exam Prep Included
  • Completion Badge
AI-3016: Develop generative AI apps with Microsoft Foundry (APL-3016) Badge

Earn Your Badge

Complete all lessons to unlock the AI-3016: Develop generative AI apps with Microsoft Foundry (APL-3016) achievement badge.

Skill Level Intermediate
Total Duration 16h 3m
AI-3016: Develop generative AI apps with Microsoft Foundry (APL-3016) Badge
Achievement Badge

AI-3016: Develop generative AI apps with Microsoft Foundry (APL-3016)

Awarded for completing the AI-3016 course — building a production-quality generative AI copilot on Microsoft Foundry end-to-end (Foundry account + project, model deployment, Microsoft Agent Framework, RAG, custom content-safety filters, evaluation, and an end-to-end capstone) and prepping for the APL-3016 Applied Skills assessment.

Course AI-3016: Develop generative AI apps by using Microsoft Foundry (APL-3016 prep)
Criteria Complete all 11 lessons of AI-3016, including the end-to-end Halcyon Assist capstone.

Skills You'll Earn

Microsoft Foundry Azure AI Microsoft Agent Framework Retrieval-Augmented Generation Responsible AI Content Safety Evaluation APL-3016

Complete all lessons in this course to earn this badge