Projects: Artificial Intelligence and Machine Learning¶

Project ideas organized by difficulty. Each project is a real, deployable artifact — not a toy exercise. Build these as portfolio pieces.

Table of Contents¶

• Beginner Projects
• Intermediate Projects
• Advanced Projects
• Expert Projects
• My Projects

Beginner Projects¶

Project 1: Titanic Survival Predictor¶

Modules required: 01, 03, 05 Estimated time: 6–10 hours

Build a binary classifier to predict survival on the Titanic. This is the classic ML "hello world" problem. Use the Kaggle Titanic dataset (freely available).

Deliverables: - Exploratory data analysis notebook (Markdown + code blocks — no .ipynb) - Feature engineering: handle missing values, encode categorical variables - Train and compare at least 3 classifiers (logistic regression, random forest, gradient boosting) - Evaluate with accuracy, precision, recall, F1, and ROC-AUC - Document the top 5 most important features

What you learn: - End-to-end supervised learning pipeline - Handling real-world messy data - Model comparison and selection

Project 2: House Price Regression¶

Modules required: 01, 03, 05 Estimated time: 8–12 hours

Predict house prices using the Ames Housing dataset. More features than Titanic — stronger focus on feature engineering.

Deliverables: - Full data exploration with visualizations (matplotlib/seaborn) - Feature engineering: polynomial features, log transforms, encoding - Train linear regression, ridge regression, and gradient boosting - Report RMSE and R² on a held-out test set - Residual analysis: where does the model fail?

Project 3: Customer Segmentation¶

Modules required: 01, 04 Estimated time: 6–8 hours

Segment customers into groups using unsupervised learning. Use a retail or e-commerce dataset (e.g., the UCI Online Retail dataset).

Deliverables: - RFM (Recency, Frequency, Monetary) feature construction - k-means clustering with elbow method for k selection - PCA visualization of clusters - Plain-language description of each customer segment and its business implications

Intermediate Projects¶

Project 4: Spam Email Classifier¶

Modules required: 03, 05, 08 (intro) Estimated time: 10–15 hours

Build a spam classifier from scratch using TF-IDF features and classical ML — no neural networks required.

Deliverables: - Text preprocessing pipeline (tokenization, stemming, stopword removal) - TF-IDF feature extraction - Naive Bayes, logistic regression, and SVM classifiers - Precision/recall analysis (explain the tradeoff in the spam context — which matters more?) - Cross-validation with stratified k-fold

Project 5: Image Classifier with Transfer Learning¶

Modules required: 07 Estimated time: 12–18 hours

Build an image classifier using a pre-trained ResNet or EfficientNet (PyTorch). Use a dataset of your choice (e.g., a 10-class subset of ImageNet, or a domain-specific dataset you care about).

Deliverables: - Data loading and augmentation pipeline using torchvision - Fine-tune a pre-trained model on your dataset - Training and validation loss/accuracy curves - Confusion matrix and per-class precision/recall - At least 5 misclassified examples with analysis of why the model failed

Project 6: Time Series Forecasting¶

Modules required: 03, 05, 07 Estimated time: 12–16 hours

Forecast a time series (e.g., stock prices, energy consumption, weather) using both classical methods (ARIMA) and ML methods (LSTM or Temporal Convolutional Network).

Deliverables: - Time series decomposition (trend, seasonality, residuals) - ARIMA baseline - LSTM or TCN implementation in PyTorch - Proper train/validation/test split (no data leakage across time) - RMSE, MAE, and MAPE metrics - Discussion of when each approach is more appropriate

Advanced Projects¶

Project 7: Text Summarization with HuggingFace¶

Modules required: 08 Estimated time: 15–20 hours

Fine-tune a sequence-to-sequence model (e.g., T5 or BART) for abstractive text summarization on a dataset of your choice.

Deliverables: - Data preprocessing and tokenization for seq2seq - Fine-tuning with HuggingFace Trainer API - ROUGE score evaluation - Qualitative comparison of pre-fine-tuning vs. post-fine-tuning outputs - Discussion of failure modes (hallucinations, truncation issues)

Project 8: ML Model Serving API¶

Modules required: 09 Estimated time: 15–20 hours

Package a trained model as a production-ready REST API with monitoring.

Deliverables: - FastAPI endpoint for inference (single predictions + batch) - Request validation with Pydantic - MLflow tracking of the model used - Logging of prediction latency and input distribution - Docker container for the API - Health check endpoint

Expert Projects¶

Project 9: RAG Question-Answering System¶

Modules required: 08, 10 Estimated time: 20–30 hours

Build a retrieval-augmented generation (RAG) system that answers questions about a document collection.

Deliverables: - Document ingestion and chunking pipeline - Embedding model (e.g., sentence-transformers) for semantic search - Vector store (e.g., FAISS or ChromaDB) - LLM integration for answer generation - Evaluation: accuracy on a set of held-out questions - Latency profiling and optimization discussion

Project 10: Fairness Audit of a Production Model¶

Modules required: 11, 03, 05 Estimated time: 15–20 hours

Take an existing ML model (e.g., the Titanic classifier from Project 1, or a credit scoring model from a public dataset) and audit it for fairness.

Deliverables: - Define protected attributes (e.g., sex, age group) - Compute group-specific metrics: accuracy, FPR, FNR - Apply SHAP to explain individual predictions - Identify at least 2 fairness violations - Propose and implement one mitigation strategy - Write a 1-page governance memo summarizing findings and recommendations

My Projects¶

Track your own projects here. Append a row when you start or complete a project.