AIDSR
Agentic & AI-assisted Data Science with R
AIDSR
Agentic & AI-assisted Data Science with RAI-assisted
Course description
AIDSR extends the MHDSR pathway into agentic and AI-assisted data science with R: using a command-line coding agent to drive a reproducible analysis and ML pipeline under version control, and using large language models for structured extraction and text classification in health research.
The course is hands-on and opinionated about guardrails: version control, deterministic and logged runs, evaluation against a gold standard, cost and reproducibility, privacy, and a clear sense of when not to use an LLM or an agent.
Learning objectives
By the end of AIDSR, participants will be able to:
- set up and drive a command-line coding agent on a version-controlled analysis project;
- use an agent to generate, execute, and review analysis code reproducibly;
- build an end-to-end, reproducible ML/analysis pipeline with agent assistance;
- design schemas and prompts for structured extraction from clinical text and documents;
- build and evaluate LLM-based text classification and annotation workflows;
- assess LLM outputs against a gold standard, and quantify agreement, cost, and error;
- apply guardrails for determinism, logging, privacy, and reproducibility;
- judge when an LLM or agent is, and is not, appropriate for a task.
Schedule
| No. | Session |
|---|---|
| 1 | Agentic workflows and CLI coding agents: what a coding agent is, setup, project structure, version control, permissions and guardrails, safe iteration |
| 2 | Code generation, execution and review with an agent: iterating on analysis code, tests, reproducible environments, reviewing and correcting agent output |
| 3 | A reproducible pipeline end to end: from raw data to report with agent assistance, under version control, with logging and re-runs |
| 4 | LLMs for structured extraction: schema design, prompting, few-shot examples, validation, evaluation against a gold standard |
| 5 | LLMs for text classification and annotation: zero/few-shot classification, prompt design, inter-rater agreement, bias and failure modes |
| 6 | Guardrails, evaluation and reproducibility: determinism, logging, cost, privacy, error analysis, when not to use an LLM, final mini-project |
Practical information
- Audience
- Researchers, PhD students, clinicians, public health professionals, data analysts
- Prerequisites
- ML4HOR (or equivalent ML experience) and BIOSTATR-level R
- Format
- Live sessions, slides, notebooks, datasets, exercises
- Schedule
- Mon/Wed/Fri 18h00–20h00 (Morocco time) · 2 weeks
- Tools
- R, RStudio, Quarto, a command-line coding agent, an LLM API
- Assessment
- Reproducible Quarto mini-project
- Next step
- METAR – Conducting Meta-Analysis with R