8 Future Trajectory and Vision
My vision for the coming years is structured around three horizons: consolidating the support ecosystem, creating a new advanced course offering, and formalizing my pedagogical research (SoTL).
8.1 Horizon 1: GPT-CDA Next Step (Phase 2, 2026-2027)
8.1.1 Why a Phase 2
Phase 1 made it possible to develop the agent, integrate it into CDA facilities, train staff, test in real conditions, and progressively deploy the tool.
Phase 2 now targets sustainability and service expansion: instead of depending on an external interface, the objective is to build an institutional in-house web platform that accesses the model via an API and enables a controlled, scalable experience integrated with CDA needs.
8.1.2 Vision: Hybrid Online Platform (AI + Teaching Assistants)
The core of Phase 2 is native integration of a hybrid online pathway:
- students use GPT-CDA in self-service mode
- during opening hours, they can be transferred to an online teaching assistant (chat)
- transfer preserves relevant history and context to make human intervention more effective
This step is also a direct response to accessibility challenges: supporting students who cannot physically come to the CDA and better covering needs beyond in-person support.
Pedagogical positioning Phase 2 does not aim to replace human support. It organizes help into two layers: handling simple or recurrent requests through the agent, then human intervention when needed, with continuity of context.
8.1.3 Operational Objectives of Phase 2
Design and deploy an in-house API-based web platform, including UL authentication, access control, logging, and administration tools.
Integrate a hybrid pathway: interaction with GPT-CDA, then transfer to an online teaching assistant with context preservation during opening hours.
Implement and document a fully remote version of the service (agent + human interaction) in one course, then extend to all courses supported by the CDA.
Evaluate acceptability and large-scale adoption conditions: preferences (human, AI, mixed), barriers, user and non-user profiles, usage trace analysis.
Produce reusable deliverables: responsible-use guide, technical documentation, implementation recommendations.
8.1.4 Implementation Plan (May 2026 to April 2027)
Phase 2 is planned from May 1, 2026 to April 30, 2027.
- Analysis and specifications (May to June 2026)
- consolidate Phase 1 feedback, clarify student pathway
- confidentiality requirements, monitoring needs, transfer modalities to teaching assistants
- In-house platform development (June to September 2026)
- authenticated web interface (UL account)
- conversational model integration via API
- access management, minimal logging, admin dashboard
- Online human support module (September to October 2026)
- teaching assistant chat with queue management
- transfer mechanism from GPT-CDA preserving relevant context
- Operational pilot and testing (October to December 2026)
- controlled deployment in one fully remote course
- student and assistant feedback, fixes, stabilization
- Expanded deployment and evaluation (January to April 2027)
- extension to all CDA-supported courses
- training and documentation
- acceptability study (survey, focus groups, usage analysis)
8.1.5 Evaluation: What Will Be Measured and Why
Phase 2 explicitly documents adoption conditions to guide deployment decisions and continuous improvement.
Indicators and typical questions:
- platform usage volume and recurrence
- rate of transfer to teaching assistants and when transfer occurs
- satisfaction, perceived utility, preferences (human, AI, mixed)
- profiles of students who adopt or do not adopt, reported barriers
8.1.6 Risk Management and Responsible Implementation
Phase 2 includes explicit attention to:
- limiting out-of-scope responses (guidance toward course concepts and methods, pedagogical framing, redirection in case of uncertainty)
- confidentiality and data minimization (clear student information, validation of retention parameters)
- academic integrity: support for process rather than generation of final answers when inappropriate
- data collection under required authorizations, including ethics process when needed
Watch point The added value of Phase 2 depends on two difficult but essential elements: governance (access, logging, confidentiality) and context transfer to humans. The in-house platform is precisely the lever that makes these elements controllable.
8.1.7 Expected Outcomes and Sustainability
Expected outcomes:
- increased accessibility to support (including for remote students)
- reduced waiting times and better support continuity
- better use of teaching assistants’ time (higher-value interventions)
- reusable model adaptable to other contexts
Sustainability:
- the in-house platform reduces dependency on an external interface and facilitates maintenance
- technical documentation and user guide support continuity beyond grant funding
8.2 Horizon 2: STT-4230 / STT-6230 - Programming for Data Science: Development of the New Course (Summer 2026)
8.2.1 Context and Pedagogical Intent
In summer 2026, I am developing a new version of STT-4230/6230 Programming for Data Science. This redesign aims to align the course with current practice: students should not only write code that runs, but learn to design, structure, test, optimize, document, and deliver reproducible artifacts in a collaborative setting. The course is explicitly designed for students who already have R experience, so class time can focus on intermediate-to-advanced competencies: project engineering, performance, software architecture, and full model lifecycle.
8.2.2 Course Vision: Moving from “Script” to “Product”
The redesign is guided by a progressive transition:
- from one-off scripts to structured projects
- from local results to shareable deliverables (reports, dashboards, web apps)
- from individual code to collaborative code (versioning, review, documentation)
- from “I can run something” to “I can deploy and maintain a pipeline”
The objective is that students leave the course able to deliver a mini data-science product: a clean repository, a clear pipeline, reproducible analyses, and the ability to diagnose, maintain, and evolve the project.
8.2.3 Modular Architecture (Planned Progression)
The course is organized in modules following a real production logic, where each block enables the next.
- Structuring data-science projects
- directory conventions, file organization, reproducibility principles
- dependency and environment management
- versioning best practices and commit strategy
- Data handling and management
- import, validation, cleaning
- robust and traceable transformations
- output structuring (clean tables, exchange formats)
- Visualization
- exploratory visualization and communication
- principles of readability, coherence, and graphical honesty
- graphics as components of a reproducible pipeline
- Reports, web apps, and dashboards
- reproducible reports (generation, parameters, templates)
- decision-oriented dashboards
- web apps for interaction and dissemination
- Object-oriented programming
- motivation in data science (stable interfaces, rich objects, methods)
- object design for analytical workflows
- user-oriented design: prints, summaries, diagnostics
- Functional programming with tidyverse and pipelines
- transformation composition
- readable, testable, reusable pipelines
- “data in, data out” principles and pure functions when possible
- Profiling, optimization, and performance management
- identify bottlenecks
- optimize without harming readability
- tradeoffs: performance, maintainability, reproducibility
- Asynchronous and parallel programming
- when parallelization is relevant
- structuring parallelizable tasks, avoiding common errors
- robustness: error handling, logs, result reproducibility
- End-to-end model pipeline and hyperparameters
- structuring full machine-learning pipelines
- train/validation/test separation
- hyperparameter tuning and data-leakage prevention
- Model deployment and production
- turning a model into a service or reusable component
- producing consumable outputs (API, report, app)
- minimum requirements: documentation, tests, versioning
- Model lifecycle
- continuous monitoring: drift, degradation, updates
- logging, monitoring, and quality maintenance
- reflection on model governance and responsibility
- Collaboration and documentation: package
- turning a project into a reusable package
- user and developer documentation
- team collaboration: issues, code review, conventions, continuous integration
8.2.4 Assessments and Deliverables (Orientation)
Assessments will be aligned with delivery competencies:
- recurring deliverables as structured, reproducible repositories
- artifact production: report, dashboard, or web app
- software components: objects, functions, pipeline, tests, documentation
- final integrative project: a complete mini-product (data -> pipeline -> deliverable -> lightweight deployment), with collaboration and traceability requirements
8.2.5 Expected Outcomes
This redesign aims to produce a clear increase in autonomy and programming maturity: less trial-and-error, more intentional design; fewer fragile scripts, more maintainable projects; and a concrete understanding of what it means to put analysis, models, and code into production in real environments.
8.3 Horizon 3: Pedagogical Research (SoTL)
I plan to formalize my approach through academic publication on university pedagogy.
STT-1100 analysis: an article documenting course design is in preparation. A longitudinal study on skill retention is also envisioned to evaluate whether students immersed in the STT-1100 narrative retain better programming habits throughout their program.
Competition club: launch of a student club for Kaggle/hackathon competitions, which will also serve as an observation site for informal and peer-based learning dynamics.
8.4 Conclusion: A Pedagogy of “Doing Together”
Across these projects, my trajectory remains guided by one conviction: teaching statistics is not the transmission of inert knowledge, but an invitation to join a living, tool-enabled, and ethical community of practice.