Dedy AI Engineer: route map and grading
The 11-lesson route, prerequisites and how challenges are graded.
The route, level by level
The course climbs from basics to expert work across four levels. Each level has a gate: aim to pass at least the Medium challenge of every lesson in a level before moving up, because later lessons reuse the evaluation set you build in Lesson 2.
Level 1 · Foundations
Understand the internal mechanics and learn to measure quality. Without these two, every optimisation that follows is guesswork.
- Lesson 1. Inside an LLM: tokens, attention and samplingFoundationsSee how a model predicts the next token, and measure it yourself.
- Lesson 2. Engineered prompts and evaluation setsFoundationsWrite prompts like specifications, then prove them with numbers.
Level 2 · Intermediate
Customise model behaviour: further training, external knowledge, and letting the model act.
- Lesson 3. Fine-tuning and LoRAIntermediateWhen to train further, and how to do it cheaply with adapters.
- Lesson 4. RAG: answering from your own documentsIntermediateChunking, embeddings, retrieval, citations and faithfulness.
- Lesson 5. Agents and tool useIntermediateLet the model call functions, loop under control and know when to stop.
Level 3 · Advanced
Standardised tool infrastructure, security, and data that is not plain text.
- Lesson 6. MCP: plugging tools into any agentAdvancedBuild your own MCP server and connect it to a client.
- Lesson 7. Safety: prompt injection and jailbreaksAdvancedDefence in depth for LLM applications.
- Lesson 8. Multimodal: images, PDFs and speechAdvancedExtract structured data from images and documents.
Level 4 · Expert
Operating at real scale: cost, latency, an engineering workflow with Claude, and a complete product.
- Lesson 9. Deployment, latency and costExpertCaching, batching, model choice, monitoring and budgets.
- Lesson 10. Claude in practice: API, tools and Claude CodeExpertUse the Claude API and Claude Code like an engineer.
- Lesson 11. Capstone: a verifiable customer-support assistantExpertPut it all together into a product with evals and a cost report.
| Level | Lessons | What you build | The question you can answer |
|---|---|---|---|
| 1 Foundations | 1, 2 | A notebook that measures tokens and probabilities; a 30-case eval set | "Is the new prompt really better, and by how much?" |
| 2 Intermediate | 3, 4, 5 | A LoRA adapter; a document Q&A system with citations; a tool-using agent | "Fine-tune, RAG or prompt? At which step does the agent fail?" |
| 3 Advanced | 6, 7, 8 | An MCP server; an injection defence layer; an invoice extractor | "Where does an attacker get in, and at which layer do we stop them?" |
| 4 Expert | 9, 10, 11 | A cost sheet; a Claude Code workflow; the capstone | "What do a thousand daily users cost, and how does it break?" |
Prerequisites
- Basic Python: functions, loops, reading and writing JSON, installing packages with
pip. - Comfort with a terminal and Git at the level of clone and commit.
- An API key from at least one model provider (for example Anthropic or OpenAI). Every lab also has a route using a small open model locally or on free Google Colab.
- No advanced maths. Wherever a formula is needed, the lesson works through it with concrete numbers.
Suggested budget for the whole course: under 10 USD of API usage if you use small models for most experiments. Lesson 9 teaches you to measure this yourself.
How each lesson is built
- Objectives: three to five things you can do afterwards.
- Short explanation: only the theory you actually need for the lab.
- Lab: step-by-step, with starter code.
- Challenges: Easy, Medium and Boss, each with acceptance criteria.
- Self-check quiz: short questions, with answers in the "Answers" section.
- Further reading: official vendor documentation and the original papers.
How challenges are graded
| Level | Points | Meaning |
|---|---|---|
| Easy | 10 | Proves you ran the lab and understood it |
| Medium | 20 | You must change the design, not just a parameter |
| Boss | 40 | An open problem with measurable results and a short written report |
A challenge scores only when it meets every acceptance criterion. There is no partial credit: this is how real product teams accept a feature.
A submission contains the source (repository or notebook), a RESULTS.md file with the command to reproduce, the numbers you measured, and a paragraph on "what I would do differently". Boss submissions must reproduce on the grader's machine with a single command.
| Total (10 lessons + capstone) | Dedy grade |
|---|---|
| ≥ 600 | Dedy AI Engineer, Distinction |
| 450 to 599 | Dedy AI Engineer |
| 300 to 449 | Course completed |
The capstone has its own 100-point scale (see Lesson 11) and is mandatory for the Dedy AI Engineer grade.
Honesty rules
- Using AI to help write code is allowed and encouraged. You must still be able to explain every line you submit.
- Never submit numbers you did not produce. Record the model, run date, temperature and seed where relevant.
- Do not put other people's real personal data into your test sets.