AI-103 Developing AI Apps and Agents on Azure Roadmap

Start with architecture and responsible operation. Select services and models through evidence, then design identity, networking, deployment, CI/CD, quota, cost, safety, and observability before adding autonomous tools.

• Map tasks to LLMs, small models, code models, multimodal models, and Foundry Tools
• Compare model quality, modality, context, latency, safety, availability, quota, and cost
• Choose Foundry services for generation, grounding, vector search, tools, memory, and workflows
• Design Foundry projects, model deployments, agent deployments, environments, and rollback
• Version code, prompts, model choices, instructions, tools, indexes, safety settings, and evaluation data
• Use Microsoft Entra ID, managed identities, keyless credentials, least-privilege RBAC, and Key Vault when a secret is unavoidable
• Plan private endpoints, private DNS, restricted public access, controlled egress, and end-to-end network tests
• Manage TPM, RPM, concurrency, scaling, retries, budgets, and per-workload cost attribution
• Configure content filters, Prompt Shields, moderation, provenance, oversight, approvals, and audit workflows
• Combine Foundry evaluation, Application Insights monitoring, and OpenTelemetry tracing

Spend the most time here. Build a cited RAG application, then turn it into an agent with narrow tools. Keep model planning separate from identity, authorization, validation, approval, and execution.

• Connect an application to a Foundry project and consume task-appropriate model deployments
• Engineer prompts, structured outputs, parameters, history, context, and measurable acceptance criteria
• Implement RAG with source metadata, citations, out-of-corpus behavior, and retrieval evaluation
• Define agent role, goals, conversation state, memory approach, tool schemas, budgets, and stopping conditions
• Compare Azure AI Search, File Search, web search, Code Interpreter, Azure Functions, OpenAPI, MCP, and custom functions
• Validate every tool argument, identity, scope, business rule, timeout, error, and idempotency key in trusted code
• Require deterministic human approval for consequential actions and retain an independent kill switch
• Design bounded multi-agent roles, handoff contracts, shared state, least-privilege tools, escalation, and cycle prevention
• Trace models, retrieval, decisions, handoffs, tools, approvals, errors, latency, and tokens
• Evaluate groundedness, relevance, safety, task adherence, completion, navigation, tool selection, inputs, outputs, and cost

Practice both focused vision APIs and multimodal models. Preserve visual evidence, review accessibility output, and treat text embedded in images as untrusted data.

• Select image or video generation and editing models based on task, controls, region, safety, and cost
• Understand prompt-driven generation, reference media, inpainting, masks, and editing workflows conceptually
• Use multimodal models for grounded questions about visual evidence
• Use Image Analysis captions, dense region captions, OCR, objects, people, and smart crops where supported
• Create concise alt text and extended descriptions and review them against accessibility needs
• Use Content Understanding for schema-defined image or video extraction and RAG-ready representations
• Compare Content Understanding, Vision, Video Indexer, and multimodal models by required output
• Moderate unsafe visual and multimodal content with appropriate controls
• Detect indirect prompt injection in image text and keep authorization outside visual instructions
• Evaluate missing objects, hallucinated objects, spatial relations, regions, language support, safety, latency, and cost

Choose between deterministic prebuilt language capabilities and generative extraction. Add voice only with representative audio, privacy controls, and end-to-end task evaluation.

• Extract named entities, key phrases, topics, summaries, and validated structured JSON
• Distinguish NER from PII detection and map redaction to organizational policy
• Use sentiment analysis and opinion mining to associate sentiment with specific aspects
• Detect language, harmful text, sensitive content, tone, and domain-specific errors
• Compare Azure Translator with LLM-powered translation for the scenario and supported languages
• Customize domain extraction or summarization only after establishing a baseline and evaluation set
• Implement real-time or batch speech to text according to interaction and latency needs
• Use phrase lists or custom speech when measured domain vocabulary or acoustic errors justify adaptation
• Implement text to speech with appropriate voices, SSML, accessibility, consent, and disclosure
• Evaluate language and speech quality, privacy, fairness, latency, cost, and human escalation

Finish with ingestion and retrieval because they connect every modality to agents. Preserve structure and provenance, evaluate failures by stage, and complete both synthetic portfolio projects.

• Ingest approved documents, images, audio, and video with validation, metadata, versioning, and failure isolation
• Use OCR, layout analysis, tables, figures, fields, Markdown, and source locations for downstream reasoning
• Choose Content Understanding or Document Intelligence by modality, structure, labels, schema, accuracy, and cost
• Use prebuilt or custom analyzers and route low-confidence or inconsistent records to human review
• Configure Azure AI Search indexes, text and vector fields, compatible embeddings, filters, and semantic configuration
• Compare keyword, vector, hybrid, filtered, semantic, and reranked retrieval on a versioned query set
• Monitor indexers, enrichment failures, extraction quality, freshness, vector consistency, index health, and relevance drift
• Connect approved retrieval pipelines to applications and agent knowledge tools with citations
• Complete the secure cited support agent and multimodal extraction projects using synthetic data
• Use timed original questions, explain every distractor, revisit weak objectives, verify current docs, and clean up labs