Follow the weeks in order or use them as modules at your own pace. Each lecture has a reading view, a canonical production failure mode to inspect, and its original slide deck.
In the taught university offerings, the course was assessed through three graded phase assignments (at the end of Phases 1, 2, and 3) followed by the capstone project (spanning Phases 4 and 5). In this public edition, use those phase assignments and decision artifacts as self-study checkpoints feeding into the Week 15 capstone. Check MCP version compatibility before the server exercises.
Foundations and reasoning
Understand model capabilities and inference-time scaling, engineer reliable structured prompts, design deterministic state machines before adding autonomy, and represent domain knowledge with embeddings and hybrid indexes.
| Week | Topic and suggested practice | Materials |
|---|---|---|
| Week 1 | The AI model landscapeReasoning models, inference-time compute, model selection, and the limits of benchmarks. Production failure mode: Confident reasoning over fabricated premises Suggested practiceCompare two available models on the same task. Record model versions, outputs, latency, and the limits of your comparison. | NotesSlides |
| Week 2 | Prompt engineering and structured outputsPrompt design, structured responses, tool use, prompt versioning, and reliability. Production failure mode: Unconstrained chain-of-thought overthinking and post-hoc rationalization Suggested practiceCreate a versioned prompt with a structured output schema and an evaluation set. Compare it with a simple baseline. | NotesSlides |
| Week 3 | The AI developer stackFramework choices, agent loops, observability, error handling, and deployment patterns. Production failure mode: Premature agent-ification and opaque framework wrappers Suggested practiceBuild a tool-using agent with two tools, structured output, and at least 10 scored evaluation cases. The lecture uses PydanticAI. | NotesSlides |
| Week 4 | Knowledge representationEmbeddings, vector search, hybrid retrieval, knowledge graphs, and context design. Production failure mode: Bi-encoder top-k similarity without metadata filtering or reranking Suggested practiceIndex a small corpus. Compare keyword, vector, and hybrid retrieval on the same hand-labeled questions. Compile the Phase 1 Baseline, Control-Flow & Indexing Spec. | NotesSlides |
Knowledge augmentation and evaluation
Architect hybrid and graph-based retrieval pipelines, build automated evaluation harnesses, and measure faithfulness before scaling complexity.
| Week | Topic and suggested practice | Materials |
|---|---|---|
| Week 5 | Retrieval-augmented generationChunking, retrieval, reranking, grounding, and iterative retrieval workflows. Production failure mode: Conflating retrieval misses with generation hallucinations Suggested practiceBuild a RAG baseline with citations. Track retrieval failures separately from answer-generation failures. | NotesSlides |
| Week 6 | GraphRAG and global contextGraph-based retrieval, community summaries, global questions, and evaluation tradeoffs. Production failure mode: Deploying GraphRAG on low-degree corpora or point-lookup workloads Suggested practiceCompare graph-based retrieval with the Week 5 baseline on both local and corpus-wide questions. | NotesSlides |
| Week 7 | Evaluation and LLM-as-a-judgeEvaluation layers, RAG metrics, judge biases, calibration, and human review. Production failure mode: Uncalibrated LLM-as-a-Judge position, verbosity, and self-preference bias Suggested practiceEvaluate two retrieval or prompt variants on 50 questions. Report faithfulness, relevance, context precision, and context recall. Produce the Phase 2 Retrieval & Eval Harness Report. | NotesSlides |
MCP and agentic systems
Standardize tool integration with the Model Context Protocol (MCP), coordinate multi-step agentic workflows, engineer failure recovery, and design human-in-the-loop escalation.
| Week | Topic and suggested practice | Materials |
|---|---|---|
| Week 8 | Model Context ProtocolTools, resources, prompts, transport, authorization, and host/server responsibilities. Production failure mode: Context bloat from inline tool dumps and unconstrained schemas Suggested practiceInspect a tool call and document its schema, authorization requirements, and failure behavior. Identify the protocol version used. | NotesSlides |
| Week 9 | Building MCP serversServer implementation, validated tool handlers, testing, remote deployment, and operational safety. Production failure mode: Swallowing business errors as transport exceptions (or vice versa) Suggested practiceImplement a small MCP server against a disposable data source. Test invalid inputs, rejected access, and real tool calls. | NotesSlides |
| Week 10 | Multi-agent workflowsCoordination patterns, shared state, tool isolation, cost controls, and failure recovery. Production failure mode: Unbounded retry loops and cascading failure across cooperating agents Suggested practiceCompare a single-agent baseline with one coordinated workflow. Measure task completion, latency, and cost. Draft the capstone proposal. | NotesSlides |
| Week 11 | Human oversight and memoryApproval gates, escalation, persistent state, memory retrieval, and privacy. Production failure mode: Uncheckpointed autonomous side effects and stale memory poisoning Suggested practiceAdd an approval gate and durable state to a workflow. Test rejected actions, restart behavior, and deletion of stored information. Complete the Phase 3 MCP Tool Contract & Recovery Trace. | NotesSlides |
Efficiency and governance
Optimize unit economics and latency via small language models (SLMs) and routing, distill or fine-tune deliberately against a baseline, and red-team adversarial failure modes.
| Week | Topic and suggested practice | Materials |
|---|---|---|
| Week 12 | Small models and model routingTask-specific benchmarking, quantization, serving, hardware limits, and model cascades. Production failure mode: Miscalibrated router confidence sending edge cases to the fast tier Suggested practiceBenchmark three model sizes on a 200-example task set and plot cost against quality. Collect the capstone's first evaluation results. | NotesSlides |
| Week 13 | Distillation and fine-tuningSynthetic data, LoRA, QLoRA, preference tuning, evaluation, and data provenance. Production failure mode: Teacher-error amplification and near-duplicate leakage in synthetic SFT data Suggested practiceDesign a small fine-tuning experiment with a held-out test set and documented data permissions. Compare with the untuned baseline. | NotesSlides |
| Week 14 | Safety and red-teamingPrompt injection, agent hijacking, guardrails, least privilege, and adversarial evaluation. Production failure mode: Indirect prompt injection hijacking privileged agent tools Suggested practiceAdd layered controls to a chatbot or agent. Test both blocked attacks and legitimate tasks; record failures, complete the Phase 4 Pareto Frontier & Red-Team Audit, and freeze the capstone demo. | NotesSlides |
Capstone
Build a reproducible, end-to-end AI system and defend its architectural decisions with quantitative evaluation evidence.
| Week | Topic and suggested practice | Materials |
|---|---|---|
| Week 15 | Capstone: build, measure, explainA working AI system, architecture writeup, measured evaluation, and an honest failure analysis. Production failure mode: Reporting aggregate scores without a categorized failure taxonomy Suggested practiceDeliver reproducible code, a 15-minute demonstration, a 4–6 page writeup, and evaluation results. Use the published capstone rubric for self-assessment. | NotesSlides |