AI-103 study plan

Provision an Azure AI Foundry project with a connected Azure OpenAI resource using the Azure CLI, and verify both key-based and keyless endpoints resolve.

Deploy one GPT-class chat model and one embedding model, then document the quota, tokens-per-minute, and regional availability tradeoffs of each.

Authenticate a Python client with a managed identity via Microsoft Entra ID and confirm no secrets are stored in code or config.

Build a streaming chat completion app that logs prompt/completion token counts per request for both deployments.

Implement a retrieval-augmented generation pipeline over roughly 30 documents using Azure AI Search with hybrid vector plus semantic ranking.

Compare two chunking strategies (fixed-size with overlap vs. layout-aware) and quantify retrieval quality against a 20-question golden set.

Author system prompts and few-shot examples that force JSON-schema-conformant output, and validate responses against the schema programmatically.

Wire up tool/function calling so the model can query an external HTTP API and feed results back into the conversation.

Create an agent with Azure AI Agent Service, attach at least three tools (knowledge retrieval, code interpreter, custom function), and run multi-turn conversations.

Implement a routing-based multi-agent orchestration where a coordinator delegates to two specialist agents.

Add guardrails: content safety thresholds, groundedness checks against retrieved sources, and a prompt-injection test set.

Instrument tracing and logging for agent runs, then score 10 scripted tasks against a pass/fail success rubric.

Extract fields from invoices and receipts using Document Intelligence prebuilt models and map results into a typed object model.

Train a custom extraction model on a labeled sample set and report per-field precision and recall on a held-out split.

Implement OCR fallback for scanned pages and layout-aware chunking for downstream retrieval.

Design a confidence-threshold routing step that escalates low-confidence extractions to human review and log the escalation rate.

Detect and redact PII in text and documents, then verify redaction coverage with a self-built leakage test set.

Build a conversational language understanding project with intents and entities, and evaluate it on a labeled utterance set.

Add sentiment with opinion mining and key phrase extraction to a feedback-analysis pipeline.

Configure text and document translation with a custom glossary and confirm term consistency across 5 sample documents.

Generate captions and answer visual questions with a multimodal model, then compare output quality and latency against the classic image analysis API on the same 20 images.

Extract printed and handwritten text with OCR and post-process the raw output into structured key-value fields.

Train and evaluate a custom image classification model on a labeled dataset, and report accuracy and confusion matrix.

Measure end-to-end latency and per-call cost for each vision path and record when to choose custom over prebuilt models.

Design a deployment topology covering project/hub structure, private endpoints, Key Vault, and least-privilege role assignments, and document it as a diagram.

Configure diagnostic settings, metric alerts for latency and error rate, and a cost budget with an alert threshold.

Implement responsible AI controls: content filter levels, abuse-monitoring and retention settings, and an evaluation gate before promotion.

Automate provisioning with Bicep or Terraform and prove repeatability by tearing down and redeploying from scratch.

Take timed practice assessments for all five domains and log the specific objectives missed.

Rebuild the two lowest-scoring objectives hands-on rather than re-reading notes.

Write a one-page decision cheat sheet mapping each domain to the Azure service and API choice triggers that distinguish similar options.

Re-run the two weakest domain assessments and confirm measurable score improvement.