AAIA study plan

Distinguish supervised, unsupervised, and reinforcement learning by writing one original example of each drawn from a business workflow you know

Explain bias-variance tradeoff and overfitting in your own words and identify which regularization technique addresses each failure mode

Compute precision, recall, F1, and accuracy by hand for a small confusion matrix and state when accuracy is misleading

Map the AWS shared responsibility model boundaries relevant to ML workloads by listing which controls sit with the customer

Compare foundation model architectures (transformer encoder, decoder, encoder-decoder) and match each to a task type such as classification, generation, or translation

Explain how tokenization, embeddings, and attention interact during a forward pass without consulting notes

Describe the training pipeline stages from pre-training through instruction tuning and RLHF, and state what each stage changes in model behavior

Identify which hyperparameters (temperature, top-p, max tokens) control which output characteristics

Differentiate RAG from fine-tuning and justify a choice for three distinct scenarios with stated constraints

Explain how vector embeddings and similarity search support retrieval, including the role of chunking strategy and embedding model choice

Describe prompt engineering techniques such as few-shot prompting, chain-of-thought, and system prompts, and predict their effect on output quality

Summarize the causes and mitigations of hallucination in generative systems

Select appropriate foundation model services and inference options for given latency, cost, and customization requirements

Design a RAG pipeline end to end, specifying ingestion, chunking, embedding storage, retrieval, and prompt assembly steps

Evaluate model outputs using both automated metrics and human review, and define acceptance thresholds before deployment

Apply prompt caching, batching, and model routing to reduce inference cost in a described workload

Implement function calling or tool use so a model can invoke an external API, and describe how results are returned into context

Compare agentic patterns (single-agent ReAct, multi-agent orchestration) against simple prompt chaining and justify when each is warranted

Tune retrieval quality by adjusting chunk size, overlap, and top-k, and explain the effect of each on answer relevance

Diagnose common production failures such as context overflow, retrieval misses, and latency spikes, and propose fixes

Classify AI risks into bias, toxicity, privacy, and intellectual property categories and map each to a mitigation technique

Explain how to measure and mitigate bias in training data and model outputs, naming at least two concrete detection approaches

Describe transparency and explainability obligations, including model cards and disclosure of AI-generated content

Apply human-in-the-loop review to a high-stakes use case and define escalation criteria

Identify AI-specific threats including prompt injection, data poisoning, model inversion, and insecure output handling, and pair each with a control

Design guardrails for input and output filtering, including PII redaction and jailbreak detection, for a customer-facing assistant

Map data governance requirements such as residency, retention, and consent to concrete architecture decisions

Describe logging, monitoring, and audit trails needed to demonstrate compliance for a deployed model

Consolidate weak areas by re-testing every domain and logging accuracy per domain

Practice scenario-based reasoning by writing full justifications for ten mixed-domain questions under time pressure

Rehearse explaining tradeoffs between cost, latency, accuracy, and safety for three end-to-end designs

Verify recall of key definitions and metrics without notes for all five domains