Reading the course
Start with the syllabus and follow the schedule. Each lecture's reading view includes its slides in sequence, a canonical production failure mode to inspect, citations, and a downloadable Markdown copy. Slide decks remain available for presentation and keyboard navigation.
Reference execution profiles: local vs. cloud
To keep your work durable across model releases and API updates, structure your exercises and capstone around interface contracts rather than vendor-specific SDKs. Your evaluation harness, state-machine graph, and MCP server code should remain 100% identical regardless of which execution profile supplies the underlying inference:
| System component | Profile A: Zero-cost local (consumer hardware) | Profile B: Frontier cloud API | Shared contract in your codebase |
|---|---|---|---|
| Inference & tool calling | Local open-weight models (7B–14B quantized via Ollama or vLLM) | Hosted frontier models (OpenAI, Anthropic, Gemini) | Typed structured-output schema (Pydantic / JSON Schema) with explicit timeout and retry bounds |
| Embeddings & reranking | Local bi-encoder + cross-encoder reranker (Hugging Face / SentenceTransformers) | Hosted embedding and reranking endpoints | Standardized retrieval interface returning ranked chunks, source metadata, and scores |
| Tool boundary | Local stdio or HTTP MCP servers against SQLite / disposable local files | Remote HTTP/SSE MCP servers with scoped OAuth / API tokens | Validated MCP tool schemas separating protocol errors (McpError) from recoverable tool results (isError: true) |
| Model routing (Phase 4) | Fast 3B tier cascading to 14B local model on low confidence | Fast/mini hosted tier cascading to frontier reasoning model | Deterministic quality-gate verifier logging tier escalation rate, latency ($p_{50}/p_{95}$), and unit cost |
When documenting your phase decision artifacts and capstone results, record which profile you ran, the exact model and quantization or API version, and the hardware or pricing tier measured.
Primary documentation
Use the original documentation when implementing an exercise. Pin the version you test and record when you accessed the documentation.
- Model Context Protocol specification and TypeScript SDK documentation.
- PydanticAI documentation.
- LangGraph documentation.
- Hugging Face Transformers and PEFT.
- Ragas documentation.
- OWASP GenAI Security Project.
Model catalogs and prices
Model tables in lectures illustrate a point in time. Select a model from the current catalog, check its availability and terms, and measure it on your own evaluation set.
- OpenAI model documentation.
- Anthropic model documentation.
- Google Gemini model documentation.
- Meta Llama model cards and licenses.
Capstone materials
The capstone guide contains the deliverables, suggested milestones, and self-assessment rubric. Select a project small enough to evaluate, establish a baseline, and use actual measurements in the demonstration and writeup.
The public repository currently provides lecture material and the viewer. Classroom-only starter repositories, datasets, private recordings, and past student submissions are not included.
Provenance and corrections
See the update notes for the current review scope and protocol-version caveats. For corrections or reuse permission, contact Sawan Ruparel.