Applied AIwith Sawan Ruparel

Update notes

Edition governance, release lineage, protocol compatibility, and publication corrections.

Public self-study edition

Edition governance: invariants vs. frontier shifts

This course separates durable systems engineering invariants (which rarely change) from model and protocol mechanics (which are revised across release cycles).

  • The Invariants (Unchanged since v1.0): Define measurable success criteria before prompting; never increase architectural complexity (RAG, multi-step agents, fine-tuning) without a baseline evaluation harness; inspect raw failure traces before tuning hyperparameters; bound every autonomous action with deterministic state machines or human judgment.
  • Rolling Canonical URLs: Course routes (/syllabus/, /schedule/, /capstone/, /lectures/...) remain canonical across releases so bookmarks and academic citations do not break.

Release v3.0 (Edition 2027.1 — Active, revised October 20, 2026)

Edition 2027.1 (v3.0) establishes the public self-study architecture around the 15-week curriculum (423 slides) and connects each curriculum phase to a cumulative engineering decision artifact:

  • Phase 1 (Foundations & Reasoning, Weeks 1–4): Emphasizes inference-time compute budgets, reasoning-token economics, deterministic structured outputs vs. unconstrained chain-of-thought, control-flow graphs, and hybrid knowledge indexing. Introduces the Phase 1 Baseline, Control-Flow & Indexing Spec artifact.
  • Phase 2 (Knowledge Augmentation & Evaluation, Weeks 5–7): Sharpens evaluation coverage around judge-model calibration, position/verbosity biases, and hybrid lexical/vector/graph retrieval tradeoffs. Introduces the Phase 2 Retrieval & Eval Harness Report artifact and a newly written representative evaluation exercise.
  • Phase 3 (MCP & Agentic Systems, Weeks 8–11): Pairs the taught Model Context Protocol (MCP) lecture decks with v1-to-v2 protocol compatibility guidance, tool-schema error recovery patterns, and bounded state machines with human-in-the-loop interrupts. Introduces the Phase 3 MCP Tool Contract & Recovery Trace artifact.
  • Phase 4 (Efficiency & Governance, Weeks 12–14): Highlights multi-tier cascade routing, task-specific Small Language Model (SLM) distillation workflows, and adversarial red-teaming. Introduces the Phase 4 Pareto Frontier & Red-Team Audit artifact.
  • Dual-profile execution guidance: Documents local open-weight vs. frontier API reference stacks so evaluation harnesses and MCP servers remain portable across consumer hardware and cloud APIs.

This release includes a targeted publication review. It does not represent a fresh verification of every model comparison, numerical claim, code sample, cited paper, or outbound URL in the lecture collection.

Historical releases and taught lineage

  • Release v2.0 (Spring 2026 — University of Connecticut, GRAD 5900): Second graduate offering; expanded the dedicated 4-week MCP and agentic architecture block, GraphRAG vs. dense retrieval economics, and inference-time scaling foundations. All 15 lecture decks and per-week downloadable Markdown bundles are published on the schedule.
  • Release v1.0 (Spring 2025 — University of Connecticut): Inaugural graduate offering establishing the evaluation-first systems curriculum and reproducible capstone rubric.

Corrections in this edition

  • Corrected cross-week previews that pointed to the wrong next lecture.
  • Published the capstone rubric and study milestones that were previously referenced as missing documents.
  • Removed claims that classroom starter projects and private capstone recordings are present in the public repository.
  • Corrected the Llama 4 Scout hardware discussion: active parameters are not the full weight footprint. Meta's release describes a single H100 with Int4 quantization, rather than a 10 GB consumer-GPU deployment.
  • Replaced the description of the Llama license as Apache-like with a reference to its actual Community License and acceptable-use terms.
  • Clarified that host tool-approval settings are implementation-specific and that filesystem roots are not a substitute for an operating-system security boundary.
  • Replaced a broad provider-permissions table for training on model outputs with a version- and agreement-specific verification checklist.

MCP version compatibility

The original MCP lectures and server examples describe the 2025 protocol generation, including its initialization handshake and @modelcontextprotocol/sdk TypeScript package. These are historical examples, not a migration to the current SDK.

The July 28, 2026 MCP release introduced a stateless core and other protocol changes. The current TypeScript SDK documentation and migration guide describe the new packages and version handling. The v1 documentation identifies the older package as its maintenance line.

When working through Weeks 8 and 9, explicitly choose a protocol and SDK generation. Use compatible, pinned packages for a historical example, or follow the current SDK's quickstart for a new implementation. Do not combine imports or lifecycle examples from different generations. Updating and executing every MCP lab against the current SDK is a separate remaining course revision.

Model and framework references

Model lineups, API examples, benchmark scores, cost estimates, context windows, and framework behavior in the original lectures may reflect the taught editions. Consult the primary documentation before reproducing an example. Measure quality, latency, memory, and cost for your own task; published headline numbers are not a substitute for that experiment.

Scope of the public materials

The public edition includes the materials actually linked from its schedule. It does not announce a new university term or provide enrollment, academic credit, staffed office hours, instructor grading, or certificates. Suggested practice and the project rubric support independent study.

Editorial references

Presentation and reading improvements

The public edition now includes a faculty review path, an illustrative evaluation exercise, an institutional adaptation brief, and a downloadable syllabus. The exercise is newly written for independent study and is not presented as a historical university assignment.

The slide viewer now renders the full Markdown syntax, including nested lists and deeper headings. It preserves content that older card and column rendering could omit. Desktop presentations use a 16:9 canvas, with dense source slides split into labeled continuation parts. Mobile slides scroll naturally at readable text sizes. Reading views include a collapsible lecture index, normalized heading levels, and contextual notes for historical assessment language.