Projects
Browse milestone-based ML builds before you sign in. You can read every brief and starter file; running and submitting need an account.
6 of these are from-scratch replications of landmark papers, graded against the numbers the paper reports.
Customer Churn Prediction
Build a production-ready churn model for a telecom subscription business
Intermediate
EDAPreprocessingFeature EngineeringLogistic Regression
Lifecycle Segmentation for a Subscription Product
Find the customer groups a marketing team can actually act on
Intermediate
K-Means ClusteringFeature ScalingPCASilhouette Analysis
The Pipeline That Broke in Production
Four things went wrong upstream and the model is not one of them
Intermediate
Data ValidationSchema EnforcementPopulation Stability IndexDistribution Drift
Support Ticket Triage With Text
Route tickets from their words, and know when not to route at all
Intermediate
Text VectorisationMulti-class ClassificationMacro AveragingTarget Leakage
PageRank, from the paper
Replicate Page & Brin (1998) on a graph with a rank sink and dangling nodes
Page, Brin, Motwani & Winograd, The PageRank Citation Ranking, Stanford InfoLab 1998
IntermediatePaper replication Graph AlgorithmsPower IterationMarkov ChainsLinear Algebra
Catalogue Re-Ranking With Cold Start
Order 25 candidates per user, including the users you have never met
Advanced
Ranking MetricsNDCGCold StartCollaborative Features
Adam, from the paper
Replicate Kingma & Ba (ICLR 2015) and beat SGD on an ill-conditioned problem
Kingma & Ba, Adam: A Method for Stochastic Optimization, ICLR 2015
IntermediatePaper replicationPro OptimizationGradient DerivationNumPyAdaptive Methods
Card Fraud Detection Under Review Capacity
Ship a fraud model that a payments team would actually deploy
Class ImbalanceInformative MissingnessPrecision-Recall AnalysisThreshold Selection
Random Forests, from the paper
Replicate Breiman (2001): CART, bagging, random subspaces and out-of-bag error
Breiman, Random Forests, Machine Learning 45(1) 2001
IntermediatePaper replicationPro Decision TreesEnsemble MethodsBootstrap ResamplingRecursion
A Credit Scorecard That Means What It Says
Build a model whose 8% actually happens 8% of the time
Advanced
Probability CalibrationInformative MissingnessClass WeightingExpected Calibration Error
Neural Network from Scratch Path
Go from EDA to a regularized MLP on a nonlinear classification boundary
EDAFeature ScalingLogistic RegressionMultilayer Perceptron
Retail Demand Forecasting
Forecast 90 days ahead across 96 store-item series without leaking the future
Advanced
Time SeriesBacktestingLeakage PreventionFeature Engineering
The EM algorithm, from the paper
Replicate Dempster, Laird & Rubin (1977) and beat k-means on elliptical clusters
Dempster, Laird & Rubin, Maximum Likelihood from Incomplete Data via the EM Algorithm, JRSS-B 1977
AdvancedPaper replicationPro Expectation MaximisationMixture ModelsLog-Space ArithmeticUnsupervised Learning
Predictive Maintenance From Sensor Sequences
Catch the failure while it is still a trend, not a reading
Sequence WindowingTemporal Feature EngineeringGroup-Aware Cross-ValidationData Leakage
XGBoost's objective, from the paper
Replicate Chen & Guestrin (KDD 2016): second-order gain and exact greedy splits
Chen & Guestrin, XGBoost: A Scalable Tree Boosting System, KDD 2016
AdvancedPaper replicationPro Gradient BoostingSecond-Order OptimizationRegularizationDecision Trees
Inventory Reordering as a Control Problem
Learn a restocking policy, then find out a good heuristic beats it
Reinforcement LearningQ-LearningPolicy SearchSimulation
Wafer Defect Classification From Scratch
Write the convolution yourself and find out what it buys you
ConvolutionMax PoolingFeature MapsWeight Sharing
Attention, from the paper
Replicate Vaswani et al. (2017) and build an induction head that needs no training
Vaswani et al., Attention Is All You Need, NeurIPS 2017
AdvancedPaper replicationPro Attention MechanismsTransformersNumPySoftmax Stability
Physics-Informed Neural Network
Solve a 1D Poisson equation from sparse data plus the PDE residual
Advanced
Scientific MLPDE ResidualsCollocationBoundary Conditions