Tutorials#
The tutorials are rendered from the checked-in Jupyter notebooks. Saved outputs are shown when included, and every page includes a download link so you can run it locally.
Historical numeric routes are compatibility redirects, not a prescribed reading order. Notebook filenames and titles are descriptive; use the category pages as the recommended learning path. The first two paths are the golden path; the remaining categories answer specific analysis and platform needs.
Every tutorial follows the same contract: start with a concrete goal, state the prerequisites, produce a named run or artifact, and finish with a quick verification step. The notebook is the executable companion; the surrounding description tells you what to look for before you open it.
Learning paths#
Choose a path first; each category page contains descriptive notebooks and direct download links. Older numeric bookmarks continue to resolve. If you are unsure where to begin, start with Getting started and continue to Core workflows.
Categories at a glance#
The cards summarize the learning path. Open a category to see its explicit notebook toctree and the complete workflow context; no notebook is intentionally listed twice on this page.
60–90 min · base + optional notebook packages · offline
Plot payment timing, distributions, fraud scenarios, difficulty, and ML-ready features before choosing a modeling or production workflow.
60–90 min · base install · offline
Generate a world, configure it, inspect lifecycle events, and understand fraud truth versus delayed labels. Produces manifests, tables, and ledger checks.
60–90 min · base; plotting optional · offline
Build point-in-time datasets, stress scenarios, and reproducible benchmarks at 1k–10k source payments.
2–3 h · ML/serving/MLflow optional
Train and evaluate models, validate serving hand-offs, promote artifacts, and measure segmented drift. Docker services enhance but do not block offline work.
60–90 min · graph + ML extras · Neo4j optional
Export provenance-rich graphs, investigate fraud with Cypher, and build temporal PyTorch Geometric features with an evaluation split.
60–90 min · Kafka extra; Docker optional
Exercise logical loss, retries, duplicates, delays, reordering, outages, and Avro compatibility without simulating physical network packets.
2–3 h · service extras + Docker optional
Repair damaged projections, resume scale runs, reconcile PostgreSQL, and verify Iceberg snapshots and observability SLOs.
2–3 h · optional ML/graph/scale extras · offline
Combine calibration, counterfactuals, campaign dynamics, temporal backtesting, and scale reconciliation into reproducible evidence packages.
The notebooks use the python3 kernel. Install the project and the optional
ML extra before running the model and benchmark tutorials:
poetry install -E ml
The published site does not execute notebooks during its build. This keeps the Pages artifact deterministic; notebook JSON validity is checked separately in CI.