21FIAP MLET Course — Data Drift Detection & Monitoring
Role: Course Author / Instructor
Executive summary
Authored an 8-class graduate course on detecting, monitoring and mitigating data and concept drift in production ML systems, for FIAP's postgraduate Machine Learning Engineering program.
- Python
- scikit-learn
- Evidently
- NannyML
- River
- Alibi-Detect
- Hugging Face Transformers
FIAP's postgraduate Machine Learning Engineering program needed a dedicated discipline covering the full lifecycle of data drift, since drift is one of the leading causes of model degradation after deployment and the topic spans statistics, embeddings and production monitoring.
- Design 8 classes progressing from data/concept drift fundamentals to statistical detection (KS test, PSI), advanced multivariate metrics (MMD, Wasserstein), drift in embeddings, mitigation strategies, continuous monitoring with Data SLOs, streaming-scale architectures, and tool integration.
- Pair each class with a concrete learning objective, key theory and a matching open-source tooling stack.
- Document architecture decisions (ADRs), business cases and diagrams connecting the statistics to real operational scenarios.
I authored the course repository end to end — eight class folders each pairing a learning objective with core theory (e.g. P(X) vs P(Y|X) shifts, the Kolmogorov-Smirnov test, PSI, MMD/Wasserstein) and a matching open-source stack (scikit-learn, Evidently, NannyML, River, Alibi-Detect, Hugging Face Transformers for embedding drift). I wrote supporting ADRs, business cases and Mermaid architecture diagrams so students connect statistical concepts to production MLOps decisions, and set up a `uv`-based reproducible environment with lint/format/test tasks for future notebook and code contributions.
Published as the core teaching material for FIAP's Machine Learning Engineering postgraduate program. The expected result is student learning and hands-on practice with drift-detection tooling, not a production deployment.
Structuring a full discipline forced me to organize scattered drift-detection knowledge — statistical tests, embedding-based methods, streaming architectures — into a coherent, progressively harder curriculum, and reinforced how much documentation (ADRs, business cases) helps students bridge theory and real-world MLOps decisions.
Architecture diagram coming soon
Drop an image at assets/img/diagrams/fiap-data-drift-course.png