Worked examples: inspect a real implementation

These cases connect a question, a reviewed source, shared code, evaluation, and a readable result. Run them from the complete source or inspect the generated report first. All examples below use actual external records.

Report links open the published reading site. If a download link does not open, use the offline access route: run the module command and open the index.html it writes.

Forecast historical bike demand

Ask what information is available when tomorrow's forecast is issued. Compare persistence and trained baselines without using future weather or target components as inputs.

Module 5 lesson · Worked forecast and residual report

Compare classification with clustering

Use real wine chemistry to compare classifiers and decision trees, then examine an unsupervised grouping. Identical chemistry profiles stay within one split. A cluster ID is not a quality rank.

Module 6 classification · Classifier evidence · Module 7 clustering · Cluster profiles

Find what SMS accuracy hides

Train a vocabulary and neural classifier on real messages, compare a lexical reference, and inspect actual false negatives. Compare the current validation-selected models with the documented historical failure to understand what accuracy can hide.

Module 9 lesson · Neural NLP evidence

Evaluate local generation and actual tuning

Compare candidate likelihood with strict generation, retain invalid outputs, and inspect validation-selected LoRA training and adapter replay. Interpret the measured results using each study's actual sample and evaluation protocol.

Module 10 generation · Generation evidence · Module 11 tuning · Adapter evidence

Connect inference to monitoring

Validate the exact input schema, exercise a local app, and inspect replay over actual held-out records. The static website links the app instructions; the Python app runs locally in the framework environment.

Module 12 setup and practice · Monitoring report

Audit decisions and bound tools

Inspect denominators and group evidence from a historic income benchmark. Then trace a model-selected plan whose tools and budget are checked outside the model. Valid tool selection can still misunderstand intent.

Module 13 fairness · Audit evidence · Module 14 bounded agents · Agent trace

Use the evidence responsibly

The examples teach transferable engineering decisions. They do not establish current operational performance, justify decisions about people, or certify a production system. Read the model-use summaries, data rights, and access routes.

Download the source or choose a module to make and explain your own bounded change.