Machine Learning & AI Engineering Foundations
Build interpretable prediction experiments, challenge their evaluation and package a tested inference contract using inspectable Python examples.
What you'll be able to do
- Define prediction-time features and evaluate with a time-aware split.
- Compare regression baselines using losses that match the intended decision.
- Interpret confusion matrices and select a threshold on validation data.
- Explain voting and clustering without confusing them with guaranteed model quality.
- Check a tiny neural model's gradient numerically.
- Build a text-vector baseline and identify its limitations.
- Version model metadata and enforce a tested inference contract.
- Deliver a reproducible equipment-turnaround decision pack with documented limitations.
Before you start
- Use Python functions, dictionaries, lists and basic file formats.
- Understand SQL joins, filtering and aggregation.
- Read basic algebra and run Python 3.12 or newer; no paid model API or cloud account is required.
The curriculum
- Freeze the Prediction Moment — Free preview
Choose the target, feature availability and time boundaries before fitting a model.
- Choose a Loss Before a More Complex Model — Free preview
Compare constant baselines to understand regression losses.
- Turn Scores into Review Decisions — Sign-in access
Use confusion counts and a stated cost model to select a threshold without confusing scores with calibrated probabilities.
- Combine Votes and Discover Groups — Sign-in access
Contrast supervised voting with unsupervised clustering using small examples whose limitations remain visible.
- Inspect a Neuron's Learning Step — Free preview
Check one small gradient numerically before relying on an automatic differentiation framework.
- Represent Text with Inspectable Vectors — Sign-in access
Build a small lexical similarity baseline and identify where word overlap stops representing meaning.
- Version and Test the Inference Contract — Sign-in access
Package model metadata, feature expectations and rejection behavior before exposing predictions to another component.
- Deliver the Repair Forecast Decision Pack — Sign-in access
Integrate the experiments into a reproducible equipment-planning project with clear human-review boundaries.
Repair Forecast
Create a local decision-support experiment for a fictional equipment repair desk. Estimate turnaround or flag cases likely to exceed an agreed planning window using only information available when a request opens. The output supports a coordinator's planning review; it does not diagnose faults, rank staff, allocate employment opportunities or make safety-critical maintenance decisions.
Deliverables
- A synthetic equipment-request dataset with a feature-availability dictionary.
- A chronological split manifest and excluded-record report.
- Regression and classification baseline results with error analysis.
- A versioned inference contract with rejected-input tests.
- A model card, reproducible evaluation command and human-review decision pack.
Review your work
- Features contain no closure-time facts or post-outcome technician notes.
- Training, validation and test periods have distinct documented purposes.
- Threshold selection never uses the final test labels.
- Inference rejects missing, non-finite and unsupported values.
- Reported results are labeled synthetic experiments, not production accuracy guarantees.