Detecting Fraudulent Sales Transactions
Build an audit queue from noisy sales records with scarce labels. Best if you already know the modeling workflow.
Explore the project Advanced practice with scarce labels.Projects
Messy datasets. Real business decisions. Choose a starting point that matches your experience.
Free Shipping Threshold Experiment
Clean an experiment export, compare outcomes, and recommend whether to change the shipping threshold.
View project & prerequisites Start here after Python and experimentation.Choose a practice area
Build from a business decision to a working system
Download the data, work the brief end to end, and deliver your analysis the way you would on the job.
Decide with evidence
Experimentation & analytics
Turn messy observations and controlled experiments into defensible product decisions.
1 projectBuild predictive systems
Machine learning
Train, validate, and communicate models across tabular, recommendation, forecasting, and vision work.
6 projects01Advanced
Machine learningDetecting Fraudulent Sales Transactions
Detect fraud in 500k reported sales when only a few hundred are confirmed fraud and 96% were never inspected — a semi-supervised problem where the fraud signal must be engineered and the uninspected majority is used, not thrown away.
Data cleaningFeature engineering
03Intermediate
Machine learningInsurance Risk Tier Assignment
Assign each incoming auto policy application to one of six ordered risk tiers at quote time -- using only information that actually exists before the policy is bound.
Ordinal classificationLeakage auditing
04Advanced
Machine learningSpare Parts Demand Forecast
Clean a messy, genuinely intermittent parts-demand history and build a MASE-evaluated forecast that beats a simple baseline.
Time-series data cleaningIntermittent-demand forecasting
05Advanced
Machine learningAd Click & Conversion Prediction
Clean a raw ad-impression log and build models that predict clicks and conversions well enough -- and honestly enough -- to power a real feed-ranking system.
Class-imbalance handlingClassification metrics (ROC-AUC / PR-AUC)
06Advanced
Machine learningMovie Recommendation Engine
Build a collaborative-filtering recommender for a streaming catalog from a sparse 1-5 ratings log -- predict ratings and rank personalized top-N picks -- evaluated with RMSE and Precision@N / Recall@N.
Collaborative filteringNeighborhood methods (user–user, item–item)
07Advanced
Machine learningHierarchical Grocery Product Classification
Classify real store-shelf photos into a three-level product taxonomy -- department, product class, variety -- well enough to power a self-checkout that suggests the right produce button.
Transfer learning & fine-tuningHierarchical classification
Build systems that use models
Agentic AI systems
Design retrieval, tools, memory, permissions, multi-agent coordination, and evaluations around language models.
3 projects08Advanced
Agentic AI systemsCompetitive Catalog Intelligence Agent
Match a retailer’s product catalog to competitor products using normalized attributes, embedding retrieval, vector search, evidence-grounded decisions, and review-ready price intelligence.
Product entity matchingEmbeddings & vector search
09Advanced
Agentic AI systemsUnified Product Taxonomy Agent
Classify inconsistent retailer listings into one in-house category tree using product text, breadcrumbs, hierarchical retrieval, evidence checks, and deliberate human review.
Hierarchical classificationTaxonomy retrieval
10Advanced
Agentic AI systemsRepository Repair Agent
Build an agent that investigates a reported defect, edits the right files, runs focused tests, and stops only when the patch satisfies the repository—not merely the prompt.
Agent–computer interfacesPlanning and tool use
Act over time
Sequential decisions
Evaluate policies whose current actions change the state and rewards that follow.
1 project