A good fit if you…
- independently clean and explore unfamiliar tabular data;
- understand descriptive statistics and analytical limitations;
- want to progress from explanation to prediction;
- will run experiments and document evidence between sessions.
Software Signal Learning · Stage 3 · In pipeline
Practical Machine Learning Foundations
Build, evaluate, and critique a reproducible predictive system through honest baselines, disciplined evaluation, and evidence-based error analysis.
This course focuses on selected classical supervised models. It does not teach deep learning, LLM application development, or production MLOps.
Designed for: learners who can independently clean, explore, visualise, and explain tabular data.
- Planned format
- Live online cohort
- Proposed teaching time
- 16 sessions × 90 minutes
- Expected level
- Data-analysis fluency
- Applied outcome
- Reproducible predictive system
Starting point
Is this course right for you?
Probably not the right fit if you…
- still need guided help with pandas and visualisation;
- want deep learning, LLMs, or production MLOps;
- want algorithms without evaluation discipline;
- expect a career or accuracy guarantee.
Prerequisites
- Reproducible Python data-analysis workflows
- Tabular cleaning, exploration, and visualisation
- Descriptive statistics and limitation-aware reporting
Check your current readiness
Can you independently clean, explore, visualise, and explain an unfamiliar dataset using a reproducible Python workflow?
What you will be able to do
- Frame useful classification and regression questions.
- Define features, targets, units, prediction time, and baselines.
- Design honest train, validation, and test splits.
- Build leakage-resistant preprocessing and model pipelines.
- Compare selected linear, neighbour, tree, forest, and boosting models.
- Select metrics from real error costs.
- Reason about imbalance, calibration, and thresholds.
- Use cross-validation and disciplined tuning.
- Interpret behaviour and analyse errors by segment.
- Document limitations and make an evidence-based use or reject decision.
Capability depends on suitable prior readiness, deliberate practice, and completed work; no career or performance outcome is guaranteed.
Applied outcome
Build, evaluate, and critique a predictive system
Train a model on a real prediction problem, then show why an apparently strong score may still be unsafe or useless.
- Frame
- Baseline
- Split honestly
- Pipeline
- Compare
- Change threshold
- Analyse segments
- Use or reject
Deliverables
- Reproducible repository and data/split assumptions
- Baseline and model-comparison evidence
- Threshold and segment-level error analysis
- Model card or limitations note with a recommendation
Boundary: a reviewable classical predictive workflow—not deep learning, GenAI, deployment infrastructure, or a headline-accuracy demo.
How the course is planned to work
Learn live
Sixteen proposed 90-minute online sessions combining explanation, demonstrations, and guided practice.
Experiment and build
Short experiments, model comparisons, error analysis, and capstone repository work between sessions.
Review evidence
Selected submission review and doubt-clearing or office-hour support are planned; exact support and recording policies remain under validation.
Current curriculum design · Four phases · Sixteen lectures
Detailed curriculum preview
The complete design is public. Teaching sequence and exercise mix may be refined before launch without changing the course goal or boundary.
Phase 1 — Framing a valid machine-learning experiment
Decide whether ML is appropriate, define the task and prepare features without leakage.
Lecture 1 — From analytical questions to machine-learning problemsDecide whether a problem needs ML and frame a useful prediction task.
Problem framing
- descriptive, diagnostic, predictive, prescriptive; classification versus regression; unit of prediction; prediction time
AI/ML foundations
- rules versus learned patterns; supervised and unsupervised learning; training versus inference
Usefulness
- decision supported; cost of errors; when a rule is better; ethical and operational constraints
Lecture 2 — Features, targets, datasets, and baselinesRepresent a modelling dataset and establish a meaningful baseline.
Dataset structure
- feature matrix, target vector, numerical/categorical features, labels, row independence
Baselines
- mean/median regressors, majority classifier, business rules
Discipline
- minimum bar, assumptions, reproducible random seeds
Lecture 3 — Training, validation, test data, and generalisationEstimate whether a model learned a reusable pattern.
Splits and generalisation
- train/validation/test roles, holdout, stratification, memorisation, over/underfitting, bias–variance intuition
Split design
- time-dependent data, grouped observations, leakage through splitting
Lecture 4 — Preprocessing numerical and categorical featuresPrepare data without contaminating evaluation.
Preparation
- missing values, scaling, outliers, one-hot encoding, ordinal caution, unknown categories
Transformers
- fit versus transform, ColumnTransformer, preprocessing inside evaluation
Phase 2 — Understanding selected model families
Build intuition for focused classical models, their assumptions and trade-offs.
Lecture 5 — Linear regression and regression thinkingBuild and interpret a regression model and its errors.
Model and implementation
- line of best fit, coefficients, intercept, residuals, loss intuition, LinearRegression, prediction
Evaluation
- MAE, MSE, RMSE, R², business meaning, residual analysis, non-linearity
Lecture 6 — Logistic regression and probability-based classificationUnderstand probabilities and threshold-based decisions.
Classification
- log-odds concept, probability output, decision threshold, LogisticRegression, coefficient direction, regularisation awareness
Evaluation preview
- confusion matrix, precision, recall, threshold trade-offs
Lecture 7 — Distance-based learning with k-nearest neighboursUnderstand similarity-based prediction and the role of representation.
k-NN
- distance, voting/averaging, k, scaling, irrelevant dimensions, computational cost
Comparison
- decision boundaries, over/underfitting through k, classification/regression variants
Lecture 8 — Decision trees and interpretable rulesUnderstand recursive splits, flexible rules, and complexity.
Trees
- splits, impurity/information gain intuition, leaves, predictions, depth, minimum samples, pruning
Interpretation
- visualisation, feature-importance caveats, rules versus generalisation
Lecture 9 — Ensemble models: random forests and boostingUnderstand why combining models can improve performance.
Ensembles
- bootstrap samples, random features, variance reduction, sequential error correction, learning rate, estimators
Trade-offs
- performance, training cost, tuning, feature-importance limitations
Phase 3 — Honest evaluation and reproducibility
Connect scores to error costs and preserve valid experiment evidence.
Lecture 10 — Evaluation for classificationSelect metrics from false-positive and false-negative costs.
Metrics
- confusion matrix, accuracy, precision, recall, F1, specificity, balanced accuracy
Curves and fit
- ROC-AUC, precision-recall, threshold selection, cost-sensitive and segment performance, no cherry-picking
Lecture 11 — Class imbalance, calibration, and decision thresholdsTurn rare-outcome scores into responsible decisions.
Imbalance and probabilities
- accuracy failure, stratification, class weights, resampling awareness, calibration, ranking versus probability
Thresholds
- review bands, abstention, human escalation, cost matrices
Lecture 12 — Pipelines, cross-validation, and reproducible experimentsEvaluate preprocessing and models together without leakage.
Pipelines and validation
- Pipeline, ColumnTransformer, fit/predict, k-fold, stratified k-fold, groups/time awareness
Evidence
- mean/variance, parameters, metrics, data versions, reproducible notebooks/scripts
Lecture 13 — Data leakage and other ways models lieRecognise leakage, proxy shortcuts, and invalid evaluation.
Leakage and shortcuts
- target leakage, contamination, future information, duplicate entities, proxies, post-outcome features, collection artefacts
Diagnosis
- suspicious scores, feature/split review, ablation, temporal validation
Phase 4 — Selection, interpretation, and final critique
Select responsibly, inspect failures, and make an evidence-backed recommendation.
Lecture 14 — Model selection and responsible tuningCompare and tune without using the test set as a tuning instrument.
Selection and search
- simple versus complex, metric choice, cross-validation, grid/random search, search discipline and budget
Final evaluation
- locked test, comparison uncertainty, practical significance, retraining
Lecture 15 — Interpretation, error analysis, and model limitationsExplain behaviour, inspect failures, and judge acceptability.
Interpretation
- coefficients, permutation importance, partial-dependence awareness, local/global explanation
Errors and limits
- false positives/negatives, segment metrics, data-quality links, hard cases, populations, uncertainty, fairness awareness, model cards
Lecture 16 — Capstone: build, evaluate, and critique a predictive systemDeliver usefulness supported by evidence, not headline accuracy.
Workflow
- framing, baseline, split, pipeline, comparison, final evaluation
Deliverables and outcome
- repository, experiment summary, error analysis, limitations; threshold/segment changes; explain why a strong model may be unsafe or useless; bridge to GenAI and engineered systems
Expected commitment and technical readiness
Proposed teaching
24 instructor-led hours. This excludes assignments, project work, and office hours.
Independent work
Experiments, comparisons, capstone repository development, error analysis, and documentation. An exact range will be published before launch.
Technical readiness
Current Python/scikit-learn-capable laptop and statistics/data-analysis fluency. Any external compute requirements will be disclosed before launch.
Instructor
Suyog Joshi
More than 20 years across software engineering, architecture, banking, payments, and complex enterprise delivery, with an emphasis on decomposition, evidence, debugging, maintainability, risk, and operational thinking.
This course uses explicit problem contracts, reproducible experiments, baseline discipline, leakage prevention, and honest limitations—not disconnected notebook demos.
About SuyogWhat still needs to be confirmed before launch
The curriculum direction is available for review. Launch operations will be announced only after the following items have been validated.
- target learner demand and entry profiles
- lecture pace and cohort size
- assignment review and support capacity
- capstone and dataset feasibility/licensing
- recording, accessibility, tools, and certificate criteria
- schedule, fee/tax, payment, application, and selection process
Confirmed dates, fees, delivery conditions and application terms will be published before applications open.
Register interestFrequently asked questions
Do I need Stage 2 first?
Equivalent reproducible Python data-analysis experience is acceptable. If you cannot independently clean, explore, visualise, and explain tabular data, start with Stage 2.
Does this teach deep learning or LLMs?
No. The current design focuses on selected classical supervised models. Generative AI belongs to Stage 4.
Which models are included?
Linear and logistic regression, k-nearest neighbours, decision trees, random forests, and boosting—chosen to teach behaviour and trade-offs, not algorithm collecting.
Will I learn evaluation and leakage prevention?
Yes. Honest splitting, baselines, metrics, imbalance, calibration, thresholds, pipelines, cross-validation, leakage, tuning, interpretation, and error analysis are central.
Is this an MLOps course?
No. It builds reproducible modelling discipline but does not promise production deployment infrastructure.
Does registering interest reserve a seat?
No. It is not an application, waitlist position, seat reservation, payment, or preferential admission.
Are dates or fees available?
No. Neither has been announced. Confirmed details will be published before applications open.
How much statistics do I need?
Introductory understanding of distributions, mean, variance, correlation, sampling, and uncertainty is expected. The course adds theory alongside modelling decisions but does not replace data-analysis preparation.
Does the course cover unsupervised learning?
Supervised classification and regression form the main workflow. Unsupervised learning may be introduced conceptually but is not the primary promise.
Why teach only a focused set of models?
A small, well-understood set supports deeper work on baselines, evaluation, leakage, thresholds, and error analysis.
Will scikit-learn pipelines be used?
Yes. Pipeline and ColumnTransformer are central to preventing contamination and making experiments reproducible.
How are classification metrics selected?
Metrics are connected to false-positive and false-negative costs, imbalance, thresholds, and the decision context.
Will class imbalance and calibration be covered?
Yes: metric choice, class weights, resampling awareness, calibration, and operational threshold decisions. Calibration asks whether predicted probabilities correspond to observed outcome frequencies.
Is hyperparameter tuning included?
Yes, using disciplined grid or random search and validation data. Tuning is not a substitute for framing and evaluation.
Is the course useful for software engineers?
Yes, especially for engineers who want to build or review predictive workflows. The data-analysis prerequisites still apply.
What should I learn after this course?
Possible paths include Generative AI Application Development, model-development or deep-learning specialisations, MLOps, and domain specialisations depending on experience and goals.
Your position in the journey
Stage 2
Applied Data Analysis with PythonStage 3
Practical Machine Learning FoundationsStage 4
Generative AI Application Development
Applied Data Analysis with PythonStage 3
Practical Machine Learning FoundationsStage 4
Generative AI Application Development
This sequence is guidance, not a mandatory purchase path. Equivalent experience can support direct entry.
View the complete learning journeyInterested in building predictive systems honestly?
Record course-specific interest without applying, paying, joining a waitlist, or reserving a seat.
Register interest