4 Pedagogical Innovation
This section presents three innovation initiatives I led to address contemporary teaching challenges: student engagement in data science, assessment in the era of generative AI, and modernization of academic support.
4.1 Immersive Approach: STT-1100 Introduction to Data Science
4.1.1 Context and Pedagogical Intent
STT-1100 was redesigned as part of the update of the undergraduate program in statistics, now the undergraduate program in statistics and data science. The goal was twofold: update the content so it reflects modern data science practice, and renew the learning experience to strengthen student motivation and engagement.
This redesign addresses a common limitation of introductory courses: progression centered on isolated techniques, often disconnected from a full workflow. STT-1100 instead provides direct entry into real practice: learning data science by doing data science, with reproducible deliverables and explicit attention to tools, collaboration, ethics, and communication.
4.1.2 A Stable and Explicit Architecture: Plan, Adventure, Challenge
The course is structured around a recurring three-part framework visible throughout the course notes:
- Learning Plan: objectives, readings, preparation
- Adventure: guided lab session, professional scenario, integrated activities
- Challenge: autonomous work, graded deliverable, demonstration of mastery
This structure is intentional: it stabilizes cognitive load (students always know how a module works) while enabling gradual growth in autonomy over the semester.
Link to course notes (Quarto website): https://aureliennicosiaulaval.github.io/STT-1100_notes_de_cours/
4.1.3 Immersive Design: Learning Through Roles and Missions
A central feature of the course is narrative immersion: students take on concrete roles (junior analyst, data journalist, data engineer, etc.) and produce deliverables for an explicit audience (client, supervisor, partner team, management).
This scenario-based design serves three pedagogical functions:
- Give immediate meaning to tools and concepts (the why comes before the how).
- Strengthen communication skills (writing for an audience, not only for the instructor).
- Reproduce realistic project dynamics (constraints, traceability, revisions, collaboration).
4.1.4 Module Overview
The course includes ten thematic modules plus a final project. Each module corresponds to a mission and a role.
| Module | Theme (course notes) | Scenario and role | Main deliverable |
|---|---|---|---|
| 1 | Diving into data science | Tool discovery, exploratory mindset | First Quarto HTML report |
| 2 | GitHub and data visualization | Antarctic mission, data scientist role | Quarto notebook + GitHub repository |
| 3 | Categories in all forms | Data journalism | Quarto article for a newspaper |
| 4 | Factors and data cleaning | Junior data engineering | Script, cleaning log, clean CSV |
| 5 | Exploring relationships between variables | Statistician | Report addressing business questions |
| 6 | Collaboration and reproducibility in GitHub | Team collaboration (rotating roles) | Cross-review using GitHub issues |
| 7 | Visualization, ethics, and security | Data ethics specialist | Critical analysis and redesign of a visualization |
| 8 | Automation and web exploration | Web extraction consultant | Scraping function and metadata |
| 9 | Prediction and bias | Data scientist | Short capsule or synthesis deliverable |
| 10 | Text and dashboard | Business analysis | Dashboard and web sharing |
Note: exact titles and associated resources (plans, adventures, challenges) are available on the course website.
4.1.5 Tools and Environment: Reproducibility as a Learning Objective
The course adopts a toolset that reflects contemporary practice:
- R and tidyverse for data manipulation, visualization, and analysis
- Quarto for reports, notebooks, and teaching material (computational literacy)
- Git and GitHub for version control, collaboration, and traceability
- A course R package grouping datasets and utility functions (uniform resource distribution)
- A course-specific AI assistant, framed by responsible-use guidelines
The originality is not only using these tools, but placing them at the core of success criteria: repository organization, code clarity, workflow consistency, and ability to produce a reproducible artifact.
4.1.6 Assessments and Progression
Assessments are aligned with the targeted skills:
- Weekly challenges: regular production of reproducible deliverables
- Exams designed in the spirit of adventures and challenges (solve tasks, do not recite procedures)
- Final project: full team-based analysis, GitHub repository, written and oral communication
The final project is designed as a first complete mission: data selection, question formulation, exploration, possible simple modeling, visualization, communication, and repository reproducibility.
4.1.7 Responsible Integration of Generative AI
The course treats AI neither as a shortcut nor as a taboo. It is integrated as one resource among others, with one central instruction: verify, understand, document.
Examples of encouraged uses:
- clarifying an error message and suggesting diagnostic leads
- rephrasing or translating technical documentation
- generating a code prototype to adapt and annotate
Examples of explicitly discouraged uses:
- delegating an entire deliverable
- accepting an answer without validation
- using AI without keeping a trace of the process
4.1.8 Scholarship of Teaching and Learning (SoTL) Approach
The course redesign includes an evaluation process: an anonymous questionnaire administered after the end of the semester to document perceptions of the learning design, motivation, engagement, and material clarity.
The intent is twofold:
- improve the course based on structured feedback
- disseminate results and pedagogical design in academic venues (presentation, article), in alignment with a SoTL approach
4.2 STT-4230 / STT-6230 - R for Scientists: Mandatory and Supervised LLM Integration
4.2.1 Context and Motivation
The arrival of language models (LLMs) has made part of traditional programming and analysis assessments obsolete, especially when only a final product (script, report, answer) is assessed and can be generated quickly. In STT-4230/6230, I chose to treat this reality as an explicit learning objective: students must learn to use an LLM in a disciplined, transparent, and reproducible way, rather than trying to bypass or hide its use.
The approach is based on a simple principle: LLM use is not assessed in itself; it is assessed as a professional competency integrated into a statistical and software workflow. Students are evaluated on the quality of this integration: design, traceability, robustness, prompt quality, risk control (hallucinations, bias, confidentiality), and ability to produce an executable artifact in a clean environment.
4.2.3 Flagship Assessment: Contextual Statistical Analysis with LLM as an R Package
The key course assessment requires students to develop an R package that extends common statistical functions to produce contextualized outputs. The package format was chosen for several reasons:
- requirement for a standard software structure (DESCRIPTION/NAMESPACE, dependencies, export)
- possibility of evaluation in a clean environment (installation, execution, demonstration)
- emphasis on design, documentation, and reproducibility
The main deliverable is a package implementing “contextual” functions (for example wrappers around base functions) and associated S3 methods.
4.2.3.1 Contextualized Functions and S3 Objects
Students implement at least two contextualized functions selected from common workflows (modeling, tests, tables, visualization). Each function returns an S3 object that stores:
- the original statistical object (e.g., lm, htest)
- a context field (free text or excerpt from dataset documentation)
- reproducibility and transparency metadata (call, formula, LLM parameters, model, provider)
4.2.3.2 A “Minimal API” to Standardize Assessment Without Limiting Creativity
A frequent challenge in software design assessment is interface variability, which increases grading cost and reduces fairness. To address this, the assignment defines a minimal interface to follow:
- required function names
- argument structure
- return object type
- expected S3 methods
Students still retain broad freedom in internal implementation, prompting strategy, additional functions, and metadata enrichment. This strategy enables coherent criterion-based grading and encourages relevant extensions.
4.2.4 LLM Integration: Technical Requirements and Safeguards
LLM integration is implemented through a dedicated function, for example ctx_llm_generate(), which is mandatory. This function combines a controlled summary of numerical results with context, then queries an LLM through a local provider (e.g., Ollama) or a remote API.
Assessment enforces explicit requirements:
- anti-hallucination instruction in the prompt (do not invent values absent from results)
- context length and relevance control (e.g., extract only relevant documentation sections)
- robust fallback mode when the LLM is unavailable (clear message, no crash)
- visible separation between numerical output and interpretive text
- explicit provider and model parameters recorded in metadata
4.2.5 Student Feedback and Continuous Improvement
An anonymous post-assessment questionnaire documents perceptions of the approach. Results indicate very positive perceptions of overall learning and strong perceived value of S3 work. The relevance of LLM-generated explanations is rated positively on average, with recurring technical difficulties related to local models (connection, latency, availability). Suggested improvements converge on providing a minimal starter template and clearer expectations on package structure and interpretation quality.
This feedback feeds an iterative logic: clarify the minimal skeleton, better frame reproducibility requirements, and provide shorter but canonical examples.
4.2.6 Outcomes and Transferability
This mandatory LLM integration in STT-4230/6230 produces three major outcomes:
- training in professional LLM use (tool of work, not oracle)
- strengthened reproducibility learning (package, clean execution, fallback modes)
- more robust assessment in the face of generative AI, because it evaluates process, design, and judgment
In the medium term, this approach also supports consolidation of student contributions toward a collective project (shared package), based on open-source practices (issues, pull requests, code review), and transforms course artifacts into a sustainable resource.
4.2.7 Academic Dissemination
This pedagogical approach is discussed in an invited contribution to the Journal of Data Science, in response to Wang et al. (2025) on challenges raised by AI-generated assignment submissions. The discussion, entitled Discussion of ‘Addressing the Challenges of AI-Generated Assignment Submissions in Education’, details how this assignment operationalizes recent literature recommendations: make AI an explicit, transparent, and verifiable learning partner.
4.3 GPT-CDA - Specialized AI Assistant for the Centre de Dépannage et d’Apprentissage (CDA)
4.3.1 Context and Pedagogical Problem
The Centre de Dépannage et d’Apprentissage (CDA) in the Faculty of Science and Engineering provides support in mathematics and statistics for about twenty courses, relying on teaching assistants who support students in conceptual understanding, problem solving, and study methods.
Since 2024, two signals have been documented:
- decreased attendance: the CDA historically received about 3,000 visits per semester; since 2024, attendance has dropped by about one third
- a change in request profile: students increasingly arrive with solutions generated by AI tools (e.g., ChatGPT) that they do not always understand and that are often outside course scope
This situation creates an opportunity for pedagogical innovation: instead of competing with public AI tools, the CDA can offer a specialized, supervised assistant aligned with course expectations and designed to preserve human support.
4.3.2 Design Vision: A Hybrid AI + Human Model
The GPT-CDA project aims to develop a specialized conversational chatbot to support students in solving mathematics and statistics problems, with two guiding principles:
- disciplinary contextualization: answers aligned with expectations of courses supported by the CDA (syllabi, theory, exercise sets, past exams, internal resources)
- hybrid support: AI provides a first pedagogical mediation, then hands off to a teaching assistant when conceptual or methodological difficulty requires it
The goal is not to replace tutoring, but to better distribute pedagogical effort: provide immediate and structured support while preserving human intervention for high-value cases.
Pedagogical stance of the tool GPT-CDA is designed as a learning mediator: explanations, detection of misunderstanding, and redirection to a teaching assistant when needed.
4.3.3 Objectives and Target Population
Main objectives:
- provide self-service access to GPT-CDA in the CDA physical space
- support understanding through detailed explanations and complementary resources
- detect situations requiring human intervention and redirect to a teaching assistant
- promote responsible AI use in learning
- stimulate student engagement and motivation, and support academic success
Target population:
- students enrolled in mathematics or statistics courses supported by the CDA
- for university courses: about 2,000 students per semester, across 32 programs and 6 faculties
- added scope: remedial courses (STT-0150, MAT-0130, MAT-0150, MAT-0250, MAT-0260) reaching a broader population on campus
4.3.4 Technopedagogical Design
4.3.4.1 1) Alignment with Courses, Not with General-Purpose AI
The core design choice is integration of CDA-relevant pedagogical materials (syllabi, content, exercises, exams, internal resources), so responses remain coherent with each course framework.
This contextualization addresses a common limitation of general assistants: they may provide mathematically correct solutions that are systematically outside the course framework, which can harm understanding and learning.
4.3.4.2 2) Detection and Redirection to a Teaching Assistant
An explicit project requirement is the assistant’s ability to detect situations where human support is needed and then redirect the student to a teaching assistant.
This anchors the system in a service logic: AI serves as a first support layer but does not try to close every case autonomously.
4.3.4.3 3) Access and Support Outside Regular Hours
A key expected impact is improved access to academic support, especially outside regular CDA hours.
4.3.5 Governance, Ethics, and Responsibility
Deployment includes usage guides and integration protocols covering:
- role and limits of the tool
- redirection toward a human
- traceability of support interactions
The intention is to support autonomy without replacing human pedagogical interaction, and to promote responsible AI use in learning.
Explicit watch point The value of GPT-CDA depends on mediation quality and governance, not only model performance. This is why the project includes guides, protocols, and monitoring and evaluation tools.
4.3.6 Deliverables
The project’s core deliverables are:
- a functional platform integrating the GPT-CDA chatbot
- a user guide for students and teaching assistants