From first prompt to agents in production
29 full lessons you read and build here, in three levels. Each level ends in a system a Zambian business would actually run: a ZRA tax assistant, a production labour-compliance assistant, and a back-office agent that reconciles mobile money. Marked against a rubric, not a certificate for turning up.
29 lessons · 3 capstones · 3 lessons open to everyone · K850 for the rest, with weekly build sessions
What is in every lesson
The full lesson, here
Problem, concepts, diagrams and a step-by-step build, written for engineers. No jumping off to other sites.
Working code
Every build step with highlighted code, the complete source files and a starter app to extend.
Quizzes that teach
A warm-up before you read and a check after, each answer explained.
A tutor that stays on topic
Ask about the lesson; Jev scores how closely each answer sticks to the lesson text.
A capstone per level
Each lesson adds one milestone to a real Zambian system, marked against a weighted rubric.
Reviews and interview practice
Paste your milestone evidence for a check-by-check review; practise real interview questions.
Not sure where to start?
Answer two questions and say what you want to build. Jev picks your level and shows how sure it is; DeepSeek turns the first lesson into a project brief for your idea.
The path
Beginner. Build a grounded assistant
Prompts, schemas, embeddings, context, retrieval and tools: the parts every LLM product is made of.
You write Python comfortably and have called an LLM API. Each lesson gives you the full theory and working code; the milestone is where you stretch.
- B1Prompt Engineering: Techniques & PatternsWrite the Tax Desk’s contract and system prompt 90 minutes8 sections · 5 quiz questionsOpen
- B2Few-Shot, Chain-of-Thought, Tree-of-ThoughtTeach it the shape of a good tax answer 45 minutes12 sections · 5 quiz questionsLocked
- B3Structured Outputs: JSON, Schema Validation, Constrained DecodingEvery answer passes a schema 90 minutes8 sections · 5 quiz questionsLocked
- B4Embeddings & Vector RepresentationsEmbed the ZRA library 75 minutes8 sections · 5 quiz questionsLocked
- B5Context Engineering: Windows, Budgets, Memory, and RetrievalBuild the context the model actually sees 90 minutes8 sections · 5 quiz questionsLocked
- B6RAG (Retrieval-Augmented Generation)Answer from the documents, cite the page 90 minutes8 sections · 5 quiz questionsLocked
- B7Function Calling & Tool UseHand the arithmetic to tools 75 minutes8 sections · 5 quiz questionsLocked
- B8The ZRA Tax DeskA tax assistant for Zambian small businesses that answers only from ZRA’s published guides, cites the page behind every answer, and hands every calculation to tested code. 20 to 30 hoursCapstone · weighted rubric · data packBrief open
Intermediate. Take it to production
Evaluation before tuning, guardrails, cost in kwacha, gateways, observability and staged rollout.
You finished the beginner level or built a RAG feature before. Expect 6 to 10 hours per lesson including the milestone.
- I1Advanced RAG (Chunking, Reranking, Hybrid Search)Hybrid retrieval with reranking and role filters 90 minutes8 sections · 5 quiz questionsOpen
- I2Evaluation & Testing LLM ApplicationsThe golden set comes before tuning 45 minutes8 sections · 5 quiz questionsLocked
- I3Guardrails, Safety & Content FilteringGuards on the way in and the way out 45 minutes8 sections · 5 quiz questionsLocked
- I4Caching, Rate Limiting & Cost OptimizationKnow the cost of every answer in kwacha 45 minutes8 sections · 5 quiz questionsLocked
- I5Prompt Caching and Context CachingDesign for a high cache hit rate 60 minutes8 sections · 5 quiz questionsLocked
- I6Building a Production LLM ApplicationShip it as a service 120 minutes8 sections · 5 quiz questionsLocked
- I7LLM Observability Stack SelectionSee every request 60 minutes7 sections · 6 quiz questionsLocked
- I8AI Gateways — LiteLLM, Portkey, Kong AI Gateway, BifrostPut a gateway in front of the models 60 minutes7 sections · 6 quiz questionsLocked
- I9Shadow Traffic, Canary Rollout, and Progressive Deployment for LLMsRelease without betting the business 60 minutes7 sections · 6 quiz questionsLocked
- I10A production labour-compliance assistantAn HR and labour-law assistant for Zambian employers, built to the production standard used for regulated-domain RAG: golden set, red team, role-based access, cost per query in kwacha, drift monitoring and a staged rollout. 30 to 40 hoursCapstone · weighted rubric · data packBrief open
Advanced. Engineer agents you can trust
Tool contracts, MCP, the agent loop, state and memory, orchestration, failure modes, injection defence, evaluation and runtime control.
You finished the intermediate level or run LLM features in production. Expect 8 to 12 hours per lesson; the data pack gives you realistic statements to work on.
- A1The Tool Interface — Why Agents Need Structured I/ODraw the line between reading and acting 45 minutes7 sectionsOpen
- A2Tool Schema Design — Naming, Descriptions, Parameter ConstraintsSchemas a model can use correctly 45 minutes7 sectionsLocked
- A3MCP Fundamentals: Stateless Requests and JSON-RPCServe the tools over MCP 55 minutes7 sections · 6 quiz questionsLocked
- A4The Agent Loop: Observe, Think, ActA loop that always stops 60 minutes8 sections · 7 quiz questionsLocked
- A5Tool Use and Function CallingTool errors the agent can recover from 60 minutes8 sections · 7 quiz questionsLocked
- A6Agent Memory — Virtual Context and Memory PagingRemember customers, forget carefully 75 minutes8 sections · 7 quiz questionsLocked
- A7Stateful Graph Orchestration — Durable Execution and CheckpointsMake the workflow a state graph 75 minutes8 sections · 7 quiz questionsLocked
- A8Orchestration Patterns: Supervisor, Swarm, HierarchicalChoose the orchestration, justify it 60 minutes8 sections · 7 quiz questionsLocked
- A9Failure Modes: Why Agents BreakBreak it on purpose 60 minutes8 sections · 7 quiz questionsLocked
- A10Prompt Injection and the PVE DefenseInstructions hidden in payment references 75 minutes8 sections · 7 quiz questionsLocked
- A11Eval-Driven Agent DevelopmentGrade trajectories, not just answers 60 minutes8 sections · 7 quiz questionsLocked
- A12Production Runtimes: Queue, Event, CronRun it for a month 60 minutes8 sections · 7 quiz questionsLocked
- A13Agent Observability: Langfuse, Phoenix, OpikOperate it from the traces 45 minutes8 sections · 7 quiz questionsLocked
- A14An observable back-office agentA finance back-office agent for a Zambian SME that reconciles mobile-money statements against invoices, prepares Smart Invoice entries for review and chases overdue customers on WhatsApp, with every money-affecting action approved by a person and every step traceable. 40 to 50 hoursCapstone · weighted rubric · data packBrief open
Learn it with people, not alone
The lessons are the path. The K850 practice is the group that walks it with you: a small group, weekly build sessions and code review, in Lusaka and remote, and a certificate with a project repo you can show employers.
Built on open work
The lesson text, code and quizzes are adapted from open-source courses, credited below under their licences. The capstones, milestones, data packs, reviews and Zambian context are ours.
- AI Engineering from ScratchRohit GhumareMIT · lesson text, code and quizzes
- Awesome LLM AppsShubham SabooApache-2.0 · starter code
- AI Engineering Interview QuestionsOutcome SchoolApache-2.0 · interview questions