Applied Data Science

Applied Data Science tests whether the synthetic-SME method can produce a rigorous graduate course in a field the designer does not teach, so the quality has to come from the design rather than from subject-matter command. It is a full 15-week, three-credit, MS-level course, combining original simulations I designed for every assessment with vetted open educational resources for its readings and models.

The learners are graduate, MS-level data-science students. As a portfolio piece, the page is written for learning designers and academic leaders judging whether AI-assisted course production can hold to real quality standards.

The course is designed backward, outcomes first, then the evidence, then instruction. A panel of AI agents drafted each module against a fixed checklist of online-quality standards; I directed the panel and own every decision.

Success is a performance, not a test score. Each of the eight modules ends in a summative simulation of a real professional task, engineered so a generative model cannot complete it for the student. The Simulations tab covers how. This is authentic, performance-based assessment (Wiggins): competence is inferred from what a student does with a real, private artifact, which preserves construct validity where a conventional exam no longer can.

Every module follows the same rhythm, so the structure never becomes the obstacle (cognitive load theory, Sweller). The course is sized to the credit-hour standard the Higher Learning Commission applies (34 CFR 600.2), 135 hours of student work across 15 weeks, about nine hours a week, with an estimated-time note on every task to support planning and self-regulated learning (Zimmerman). Instructor presence is designed in from a Getting Started module and a weekly, instructor-initiated interaction plan, building the teaching presence a Community of Inquiry depends on (Garrison, Anderson & Archer).

The learning arc, week by week:

Universal Design for Learning (CAST) and WCAG 2.1 AA are built into the design rather than audited on afterward. All three UDL principles are addressed: multiple means of engagement, of representation (recorded introductions, readings, and interactive explorables), and of action and expression (the simulations and defended project let students demonstrate mastery in more than one way). Videos are captioned with slide alternatives, and time-on-task is signaled in words and a bold label, not styling alone (WCAG 1.3.1). The Getting Started module orients every learner to the online environment before content begins.

• Build and orchestration. A reusable Claude skill (the Gemini equivalent is a Gem) orchestrates the agent panel. Built with the synthetic-SME method (AI-assisted course design →).
• Design frameworks. Backward design (Wiggins & McTighe), constructive alignment (Biggs), and the revised Bloom’s taxonomy for outcome verbs.
• Quality and compliance standards, enforced on every pass. OSCQR, Regular and Substantive Interaction (34 CFR 600.2), Universal Design for Learning (CAST), and WCAG 2.1 AA.
• Open educational resources and explorables I found and license-verified. MLU-Explain, TensorFlow Playground, R2D3, Setosa, Distill, Seeing Theory, Facets, Data-to-Viz, the FT Visual Vocabulary, StatKey, Kaggle Data Cleaning, missingno, Select Star SQL, and Google’s PAIR fairness tools, Aequitas, and the What-If Tool.
• Simulations I built for this course. All eight professional simulations are my own designs, one per module. Four of them (modules 5 to 8) are built as playable interactive simulations, and I released those as open educational resources under a Creative Commons license. Made with Synthetic Data Vault and Faker for per-student seeded datasets, Evidently for data-drift detection, and Docker for reproducible environments.
• Course-grounded AI tutor. A public Gemini Notebook that answers only from the course and its vetted sources, with citations.
• Language and delivery. Python throughout; the course itself delivered as static HTML and CSS.
• Instructor materials. Eight slide decks with speaker-note scripts and a full instructor preparation kit.

Reviewed, but not yet verified by the faculty member who would teach it, not tested with students, and not delivered to students.

This course was an experiment: how far does AI get when no subject-matter expert is in the room? I do not recommend building a course that way. After it was built, a business and data analyst with an M.S. in applied statistics and data science reviewed it.

Where this is useful: as a starting point when an expert is not available yet. A faculty member handed a course on short notice, with nothing in place, gets a reviewed draft to correct instead of a blank shell. The expert still finishes it.

Download the whole course as a Common Cartridge and import it into Canvas, Blackboard, Moodle, or D2L (in Canvas: Settings, then Import Course Content, then Common Cartridge). It brings in a Start Here page, the Getting Started module, and all eight modules as pages, with the four simulations and the defense tool linked, and an editable PowerPoint lecture deck bundled inside each module. Free to reuse and adapt for noncommercial purposes under the license below.

Download the course package (.imscc)