Free to read ยท Three levels

Generative AI: evidence, tools, and accountable systems

A complete path through generation, tokens, context, prompts, structured outputs, embeddings, retrieval, RAG, evaluation, tools, agents, privacy, injection boundaries, reliability, cost, and governance.

18 lessonsEnglish + ArabicContent reviewed
Start learning

In this course

Study eighteen developed lessons in three ordered levels. Each connects a concrete application question to explicit task, evidence, permission, data, and uncertainty boundaries, with two explained examples, three diagnostic mistakes, two worked exercises, objectives, takeaways, and primary references.

The course is vendor-neutral. Hand-authored logits, vectors, labels, documents, and costs are clearly synthetic. Runnable Python examples validate small offline algorithms only; prompts, model-facing pseudocode, schemas, and review records remain non-runnable. No API charge, model or weight download, real credential, service, network call, deployment, scanner, or load experiment is involved.

What you will learn

  • Explain what generation and retrieval can establish and what still needs independent evidence.
  • Build testable input, output, tool, privacy, and stopping contracts using safe fixtures.
  • Evaluate quality, uncertainty, security, cost, and release decisions without fabricating model results.

Your learning path

Three levels

Basic: generation and evidence foundations

Separate models, tokens, prompts, structured form, vector similarity, and factual support before building an application.

  1. Models, generation, and system boundaries
  2. Tokens, context, and information budgets
  3. Prompts as task contracts
  4. Structured outputs and validation
  5. Embeddings, similarity, and index identity
  6. Factuality, uncertainty, and abstention
Explore this level

Intermediate: retrieval and controlled workflows

Connect authorized retrieval, RAG, evaluation, tools, agents, and privacy into explicit application boundaries.

  1. Retrieval, chunking, and access scope
  2. RAG, grounding, and citations
  3. Evaluation fixtures and error analysis
  4. Tool proposals and permission boundaries
  5. Agents, control flow, and stopping
  6. Privacy and untrusted instructions
Explore this level

Advanced: evaluation, reliability, and accountability

Maintain data lineage, choose interventions, budget operations, monitor change, test boundaries, and govern release and correction.

  1. Dataset lineage and governance
  2. Adaptation and model selection
  3. Latency, cost, and reliability budgets
  4. Observability, drift, and feedback
  5. Adversarial and boundary evaluation
  6. Human oversight, release, and accountability
Explore this level

Who this is for

  • Developers and technical learners designing evidence-aware generative applications.
  • Reviewers and collaborators who need precise evaluation, privacy, tool, and governance models.

Before you start

  • Basic Python, JSON, arithmetic, and application concepts; no machine-learning framework or accelerator is required.
  • No hosted-model account, API key, external dataset, model download, paid subscription, or configured service is required.

Related technical solutions

RAG answer is wrong: trace evidence before changing the promptLocate corpus, access, retrieval, packing, and claim-support failures with safe synthetic traces rather than assuming every error is a model problem.Structured candidate fails: separate parsing, schema, and supportDiagnose malformed or misleading structured output without executing it, silently inventing facts, or retrying without a total budget.Untrusted source changes the task: protect tool and data boundariesDiagnose instruction-like source content and unauthorized proposals with deterministic policy checks, safe evidence, and no real tool execution.