AI Engineering Academy

Learning objectives

Study method

1. Read and explain in your own words.
2. Reproduce examples and complete the lab.
3. Take the quiz and study every explanation.
4. Tick daily tasks only after doing the work.

Common failure modes

AI tools & developer toolkit

Tools, prices, quotas and terms change. Verify current documentation before relying on a service.

1 · Learn & reason

ChatGPT, Claude, Gemini — ask for explanations at multiple levels, worked examples, quizzes, alternative designs and critique. Ask the model to state assumptions and uncertainty. For consequential facts, verify against official documentation or a reproducible test.

Effective prompt: “Teach me [concept] from first principles. Give a worked example, one misconception, one exercise, then wait for my attempt. Do not reveal the answer until I respond.”

2 · Code & review

GitHub Copilot, Claude Code, Gemini/other coding agents — use only where available and appropriate. Ask for a small, scoped change, inspect the diff, understand each changed file, run tests, review dependencies and commit only verified work. Never give an agent unrestricted access to secrets or production systems.

Effective workflow: explain repository → request plan → approve small change → inspect diff → run tests → review security → commit.

3 · Run experiments from a phone

Google Colab, Kaggle Notebooks, browser IDEs — useful for Python, data analysis and occasional accelerator access. Save work to a repository or export files; notebook sessions can expire and free compute is limited. For backend work, use a cloud IDE when available, or practice with smaller scripts and tests.

4 · Build with models

Model APIs — compare OpenAI, Anthropic, Google and open-weight options for the exact task. Measure quality, latency, price, context limits, privacy terms and availability. Keep keys server-side, validate output schemas and design a provider fallback only after testing it.

5 · Search & knowledge

Official docs, Hugging Face, scikit-learn, PyTorch, PostgreSQL — prefer primary docs for behavior and current APIs. Hugging Face helps discover models and datasets; inspect licenses, model cards, benchmarks and compute requirements. Do not assume a popular model is suitable for your use case.

6 · Evaluate, trace & deploy

pytest, GitHub Actions, MLflow, Promptfoo, Ragas, OpenTelemetry, Docker — tests catch regressions; evaluation sets measure AI quality; traces show what happened; containers improve reproducibility. Adopt tools when they solve a real problem rather than installing every framework.

Achievements

Mastery rules

A phase is marked mastered at 80% or above. Below 80%, review explanations and complete the lab before retrying. Task completion and demonstrated mastery are tracked separately.

XP is motivational, not a credential. Be able to explain and apply the idea without assistance.