Forward Deployed Engineer Training

You can explain core artificial intelligence concepts, identify common workloads, and provision your first AI resources in Azure.

1. What AI, machine learning, generative AI and agentic AI are30 min
   • machine learning vs rule-based systems
   • large language model foundations
   • autonomous agent loops
   • neural network training paradigms
   • supervised vs unsupervised workflows
2. Common workloads: text analysis, speech, computer vision and information extraction30 min
   • natural language processing use cases
   • speech-to-text scenarios
   • computer vision pipelines
   • document intelligence extraction
   • multimodal data ingestion
3. Matching business scenarios to the right AI workload30 min
   • customer support automation
   • document processing workflows
   • multimodal data ingestion
   • sentiment analysis integration
   • predictive maintenance modeling
4. Azure subscriptions, resource groups and AI resources30 min
   • Azure portal navigation
   • resource group creation
   • pricing tiers and quotas
   • subscription cost management
   • role assignment basics
5. A guided tour of Microsoft Foundry and Foundry projects30 min
   • Foundry portal dashboard
   • creating a new project hub
   • managing connected resources
   • model catalog discovery
   • workspace permission settings
6. Foundry portal dashboard and environment configuration30 min
   • portal layout navigation
   • workspace settings configuration
   • endpoint URL retrieval
   • API key management
   • connectivity verification

Hands-on project: Provision an Azure AI resource group, create an AI service endpoint in the Azure Portal, and document your configuration in a team README.

You can evaluate AI solutions against Microsoft's six responsible AI principles and configure content safety filters in Foundry.

1. Fairness in AI solutions30 min
   • bias detection
   • dataset representation
   • mitigating disparate impact
   • demographic parity checks
   • fairness assessment tooling
2. Reliability and safety30 min
   • hallucination risks
   • model testing strategies
   • fallback mechanisms
   • robustness validation
   • failure mode analysis
3. Privacy and security30 min
   • data encryption in transit
   • Microsoft Entra ID access control
   • protecting sensitive customer data
   • credential isolation
   • secure storage accounts
4. Inclusiveness30 min
   • accessibility standards
   • multilingual support
   • user-centric design
   • assistive technology compatibility
   • cultural sensitivity evaluation
5. Transparency and accountability30 min
   • explaining model decisions
   • human-in-the-loop workflows
   • auditing system logs
   • model card generation
   • traceability standards
6. Content safety filters and guardrails in Foundry30 min
   • severity thresholds
   • blocking harmful user prompts
   • filtering generated completions
   • hate speech detection
   • sexual violence filters

Hands-on project: Configure content safety guardrails in Microsoft Foundry and test a model's response to prompt injection attacks.

You can select appropriate foundation models from the model catalog, configure endpoints, and tune generation parameters.

1. How generative AI models work: tokens, embeddings and transformers30 min
   • tokenization process
   • vector embeddings
   • transformer architecture basics
   • attention mechanism overview
   • context window limitations
2. Choosing a model from the model catalog based on its capabilities30 min
   • GPT-4o vs embedding models
   • multimodal model selection
   • benchmark evaluation
   • parameter size trade-offs
   • latency vs accuracy metrics
3. Model deployment options, quotas and endpoints30 min
   • provisioned throughput
   • pay-as-you-go endpoints
   • rate limit management
   • TPM allocation
   • scaling configuration
4. Configuration parameters: temperature, top-p and maximum tokens30 min
   • controlling output randomness
   • nucleus sampling
   • managing completion length
   • presence penalty settings
   • frequency penalty tuning
5. Comparing and evaluating models before deployment30 min
   • Foundry side-by-side evaluation
   • latency benchmarks
   • cost estimation
   • output quality scoring
   • regression testing
6. Managing rate limits and quota allocations30 min
   • tokens per minute tracking
   • requests per minute limits
   • automatic retry configuration
   • burst capacity management
   • quota increase requests

Hands-on project: Deploy a model from the Foundry model catalog and compare output variance by adjusting temperature and maximum tokens.

You can write basic Python scripts, install SDK packages via pip, and authenticate to Azure services using environment variables.

1. Python syntax refresher: variables, functions, lists and dictionaries30 min
   • basic data types
   • function definitions
   • iterating over collections
   • dictionary parsing
   • exception handling syntax
2. Virtual environments and installing packages with pip30 min
   • venv creation
   • requirements.txt management
   • pip install azure-ai-inference
   • package version pinning
   • troubleshooting dependency conflicts
3. Authenticating to Azure with the Azure CLI and Microsoft Entra ID30 min
   • az login command
   • DefaultAzureCredential
   • managing service principal secrets
   • token caching mechanics
   • RBAC role verification
4. Working with JSON requests and responses30 min
   • json module parsing
   • handling nested dictionaries
   • error catching with try-except
   • serialization formatting
   • payload validation
5. Reading and understanding SDK sample code30 min
   • navigating Microsoft documentation
   • adapting quickstart scripts
   • debugging import errors
   • type hinting in Python
   • asyncio usage basics
6. Debugging script errors and network timeouts30 min
   • traceback analysis
   • logging configuration
   • HTTP status code handling
   • network timeout configuration
   • breakpoint debugging in VS Code

Hands-on project: Write a Python script that authenticates via the Azure CLI, calls a deployed model endpoint, and parses the JSON response.

You can design effective system prompts, test them in the Foundry playground, and build a conversational chat client with conversation history.

1. System prompts versus user prompts30 min
   • defining persona and constraints
   • user input formatting
   • role separation in API calls
   • developer message configuration
   • prompt hierarchy design
2. Prompt techniques: examples, grounding and output formats30 min
   • few-shot prompting
   • retrieval grounding context
   • enforcing JSON output schemas
   • chain-of-thought prompting
   • markdown output formatting
3. Deploying a model and testing it in the Foundry portal playground30 min
   • chat playground interface
   • adjusting parameters live
   • inspecting raw API payloads
   • prompt version comparison
   • exporting code snippets
4. Building a lightweight chat client with the Foundry SDK30 min
   • streaming chat completions
   • handling message loops
   • client connection handling
   • response chunk parsing
   • asynchronous generation loops
5. Managing conversation history and token limits30 min
   • truncating message lists
   • token counting utilities
   • context window management
   • summarizing older dialogue
   • state persistence strategies
6. Handling streaming completions and error states30 min
   • delta chunk handling
   • connection drop recovery
   • rate limit exception trapping
   • graceful UI degradation
   • client-side timeout management

Hands-on project: Build a command-line chat application in Python that maintains conversation history and respects token window limits.

You can create single-agent solutions in Microsoft Foundry, equip them with built-in tools, and invoke them programmatically.

1. What an AI agent is: instructions, tools and knowledge30 min
   • agent reasoning loop
   • tool calling mechanics
   • knowledge grounding files
   • autonomous execution patterns
   • state machine transitions
2. Creating and testing a single-agent solution in the Foundry portal30 min
   • agent configuration panel
   • assigning system instructions
   • interactive chat testing
   • tool attachment workflows
   • behavior validation
3. Adding tools such as file search and code interpreter30 min
   • vector store attachment
   • code execution sandboxing
   • tool output validation
   • file upload procedures
   • execution security limits
4. Building a lightweight client application for an agent30 min
   • threads and runs API
   • polling for agent completion
   • retrieving assistant messages
   • thread message creation
   • run status monitoring
5. Testing, tracing and keeping agent behaviour safe30 min
   • tracing agent tool calls
   • preventing infinite loops
   • guarding against prompt injection
   • output sanitization
   • execution auditing
6. Monitoring agent thread runs and execution state30 min
   • run step inspection
   • failure state handling
   • thread cleanup routines
   • execution timing metrics
   • token usage auditing

Hands-on project: Deploy an AI agent with file search capabilities and build a client application that queries internal project documentation.

You can implement text analysis pipelines and integrate Azure Speech services for voice-enabled applications.

1. Text analysis: keyword extraction, entity detection, sentiment and summarisation30 min
   • named entity recognition
   • sentiment scoring
   • extractive summarization
   • key phrase extraction
   • language detection
2. Building a lightweight text analysis application30 min
   • calling text analytics endpoints
   • parsing sentiment results
   • displaying structured outputs
   • batch text processing
   • error handling for invalid text
3. Speech recognition and speech synthesis concepts30 min
   • audio sampling rates
   • neural text-to-speech voices
   • speech translation basics
   • acoustic model fundamentals
   • phoneme mapping
4. Building an application with Azure Speech in Foundry Tools30 min
   • speech SDK integration
   • audio file input handling
   • synthesizing speech output files
   • microphone stream capture
   • voice configuration options
5. Responding to spoken prompts with a deployed multimodal model30 min
   • audio input tokenization
   • end-to-end voice loop
   • latency optimization
   • streaming audio responses
   • real-time conversation flow
6. Audio format handling and transcription validation30 min
   • WAV and MP3 conversion
   • channels and bitrate settings
   • confidence score evaluation
   • punctuation restoration
   • speaker diarization basics

Hands-on project: Create a Python script that transcribes an audio file using Azure Speech and summarizes the transcript with a language model.

You can process visual data using computer vision APIs, interpret images with multimodal models, and generate images responsibly.

1. Computer vision concepts: image classification, object detection and OCR30 min
   • image tagging algorithms
   • bounding box coordinates
   • optical character recognition
   • spatial analysis basics
   • feature extraction
2. Interpreting visual input in prompts with a multimodal model30 min
   • base64 image encoding
   • multimodal prompt structure
   • visual QA workflows
   • image resolution scaling
   • multi-image context handling
3. Creating new images with generative models30 min
   • prompt engineering for images
   • aspect ratio configuration
   • image generation endpoints
   • style conditioning parameters
   • revision controls
4. Building a lightweight application with vision capabilities30 min
   • image upload handling
   • calling vision inference API
   • rendering analysis results
   • local file buffering
   • payload size optimization
5. Responsible use of image-generation models30 min
   • watermarking requirements
   • preventing abusive imagery
   • copyright considerations
   • content filtering for images
   • provenance tracking
6. Analyzing document imagery and extracting text layouts30 min
   • document layout analysis
   • reading order detection
   • handwriting recognition
   • table structure extraction
   • bounding polygon parsing

Hands-on project: Build a vision client that analyzes uploaded images for tags and object bounding boxes, passing descriptions to a chat model.

You can build structured information extraction pipelines across documents, forms, images, audio, and video using Content Understanding.

1. Extracting information from documents and forms30 min
   • form recognizer models
   • key-value pair extraction
   • table data parsing
   • prebuilt model selection
   • template-free extraction
2. Extracting information from images30 min
   • receipt data extraction
   • bounding box validation
   • confidence score thresholds
   • field normalization
   • image preprocessing
3. Extracting information from audio and video30 min
   • video indexing concepts
   • transcript entity extraction
   • timestamp alignment
   • scene change detection
   • audio stream parsing
4. Analyzers and structured output schemas30 min
   • defining custom schemas
   • JSON schema validation
   • analyzer deployment settings
   • field type mapping
   • schema versioning
5. Building a lightweight information extraction application30 min
   • calling content analyzers via SDK
   • processing extraction results
   • exporting data to CSV
   • handling asynchronous operations
   • error logging for failed extractions
6. Validating extracted data against business rules30 min
   • schema compliance checks
   • data type validation
   • range checking for extracted fields
   • human review flagging
   • exception routing

Hands-on project: Configure an Azure Content Understanding analyzer to extract structured line items from scanned vendor invoices.

You can secure Azure AI deployments using Microsoft Entra ID role-based access control and manage API keys securely.

1. Role-Based Access Control in Azure AI30 min
   • Cognitive Services Contributor role
   • assigning permissions to users
   • least privilege access principles
   • custom role definitions
   • scope inheritance
2. Managing keys and connection strings securely30 min
   • Azure Key Vault integration
   • rotating access keys
   • environment variable isolation
   • secret retrieval in code
   • access policy configuration
3. Managed identities for cloud applications30 min
   • system-assigned vs user-assigned identities
   • token acquisition without credentials
   • authenticating SDK clients securely
   • resource permission assignment
   • identity lifecycle management
4. Network isolation and private endpoint configuration30 min
   • virtual network integration
   • private link setup
   • disabling public network access
   • firewall rule configuration
   • subnet routing rules
5. Auditing access logs and security events30 min
   • Azure activity log monitoring
   • diagnostic setting export
   • security alert configuration
   • failed authentication tracking
   • compliance reporting
6. Secure credential rotation procedures30 min
   • zero-downtime key rotation
   • secondary key utilization
   • automated secret updates
   • expiration notifications
   • rollback planning

Hands-on project: Configure managed identities and secure API key rotation for an Azure AI endpoint in a resource group.

You can set up application monitoring, track token usage metrics, and diagnose deployment latency issues.

1. Monitoring resource health and metrics30 min
   • Azure Monitor dashboards
   • tracking request volume
   • monitoring error rates
   • latency percentile tracking
   • health probe configuration
2. Logging and tracing model requests30 min
   • Application Insights integration
   • inspecting payload logs
   • debugging failed inference calls
   • correlation ID tracking
   • custom telemetry writing
3. Cost management and quota optimization30 min
   • setting up budget alerts
   • managing rate limits
   • optimizing tier selections
   • cost analysis reporting
   • tagging resources for cost allocation
4. Setting up alert rules and action groups30 min
   • metric alert criteria
   • action group notifications
   • email and webhook triggers
   • severity level configuration
   • alert resolution monitoring
5. Analyzing latency and throughput bottlenecks30 min
   • time-to-first-token metrics
   • throughput saturation analysis
   • region latency comparison
   • client-side network profiling
   • payload size impact assessment
6. Diagnostic logging and log query analysis30 min
   • Kusto query language basics
   • querying request logs
   • filtering error patterns
   • exporting log analytics data
   • retention policy configuration

Hands-on project: Configure Azure Monitor alerts for token consumption thresholds and review latency metrics in Application Insights.

You can size an Azure AI deployment so it neither throttles users nor overspends.

1. What you are actually billed for20 min
   • Prompt and completion tokens
   • Per-model pricing
   • Embedding and image costs
   • Storage and ingestion fees
   • API call metering rules
2. Quotas, throttling and capacity20 min
   • TPM and RPM limits
   • 429 handling with backoff
   • Provisioned throughput units
   • Burst capacity allocation
   • Capacity planning formulas
3. Reducing spend20 min
   • Prompt and context trimming
   • Caching repeated answers
   • Routing to a smaller model
   • Batch inference pricing
   • Token reduction techniques
4. Budgets and governance20 min
   • Cost alerts
   • Per-environment quotas
   • Attributing spend to a team
   • Cost anomaly detection
   • Spending limit enforcement
5. Provisioned throughput planning and reservation models20 min
   • Reserved capacity units
   • Workload volatility analysis
   • Peak load sizing calculations
   • Cost-benefit trade-off evaluation
   • Commitment term management
6. Automated budget enforcement and cost alerting20 min
   • Budget action groups
   • Automatic scaling suppression
   • Forecasted spend tracking
   • Departmental chargeback reports
   • Cost optimization recommendations

Hands-on project: Model the token cost of the capstone workload, choose between pay-as-you-go and provisioned throughput with the arithmetic shown, implement retry-with-backoff on throttling, and add a budget alert with a documented response.

You can build and present a multimodal AI assistant, review all AI-901 exam domains, and pass timed practice assessments.

1. Capstone: a multimodal assistant combining chat, speech, vision and extraction40 min
   • architecture integration
   • end-to-end testing workflow
   • code cleanup and documentation
   • error handling orchestration
   • final presentation preparation
2. Walkthrough of every skill in the official AI-901 outline40 min
   • skill domain review
   • identifying exam trap questions
   • key terminology checklist
   • implementation weight distribution
   • core concept reinforcement
3. Microsoft practice assessment and timed mock tests40 min
   • taking timed practice exams
   • reviewing incorrect answers
   • time management strategies
   • question analysis techniques
   • score tracking
4. Exam booking, the exam sandbox and scoring40 min
   • Pearson VUE registration
   • exam environment guidelines
   • understanding the 700 passing score
   • identification requirements
   • remote proctoring rules
5. Next steps, resume, LinkedIn and mock interviews40 min
   • updating LinkedIn with certification
   • showcasing capstone projects
   • interview preparation for AI roles
   • resume bullet point refinement
   • career path planning
6. Final Q&A and certification readiness review40 min
   • doubt clearing session
   • last-minute revision topics
   • exam day checklist
   • mental preparation
   • post-exam career guidance

Hands-on project: Build and test a capstone multimodal assistant combining chat, speech, vision, and document extraction, complete with a clean shutdown script.