Learn by building
Don’t just read it — build it.
You learn AI by making things, breaking them, and fixing them. Here are 30 projects, from your first hour of Python to shipping a full AI product. Each one maps to the lesson that teaches what you need — build them as you go.
See the full roadmap →Start here — beginner projects
Small, finishable projects that lock in the basics. Build them as you reach each lesson.
Stats-from-scratch CLI
Read a list of numbers from the user and report the mean, min, and max — no libraries, just Python.
Builds on: Python Basics →Tic-tac-toe AI (minimax)
Code an unbeatable opponent with the minimax algorithm — the "classic AI" that predates machine learning.
Builds on: What AI Actually Is →Dataset detective
Load a messy CSV with pandas, answer three real questions with groupby, and back each up with a chart.
Builds on: Pandas →Coin-flip probability simulator
Simulate thousands of trials to see the law of large numbers and p-values with your own eyes.
Builds on: Probability & Statistics →Gradient descent by hand
Fit a line to points by nudging weights downhill — watch a model actually learn, no library doing the magic.
Builds on: Calculus →Iris flower classifier
The "hello world" of ML: classify flowers and plot the decision boundary.
Builds on: Classification →House-price estimator
Predict prices with linear regression and interpret the coefficients — which features actually matter?
Builds on: Regression →Handwritten-digit recognizer
Train a small neural network to read MNIST digits — your first taste of deep learning that works.
Builds on: Neural Networks →
Level up — intermediate projects
Bigger, portfolio-worthy builds that combine several skills end to end.
Spam / sentiment classifier
Turn raw text into features and classify it as spam-or-not or positive-or-negative. Measure it honestly.
Builds on: Model Evaluation →Customer segmentation
Cluster customers with k-means, then describe each segment in plain English for a "marketing team".
Builds on: Clustering & PCA →Kaggle-style boosted model
Beat your baseline with gradient boosting (XGBoost/LightGBM) and explain its top features.
Builds on: Trees & Ensembles →Image classifier w/ transfer learning
Fine-tune a pretrained CNN on your own images — a few hundred photos is enough for real accuracy.
Builds on: CNNs & Vision →Face detection app
Detect and box faces in photos or a webcam feed — a satisfying, visual first computer-vision build.
Builds on: CNNs & Vision →Customer churn, end to end
Clean → engineer features → model → evaluate, then ship it as a public repo with a proper README.
Builds on: First ML Project →A/B test analyzer
Take two variants and decide, with real statistics, whether the difference is signal or noise.
Builds on: Probability & Statistics →Semantic search engine
Search 50+ documents by meaning, not keywords, using embeddings and cosine similarity in NumPy.
Builds on: Embeddings →Recommender system
Recommend movies, books, or products by similarity — the engine behind every "you might like".
Builds on: Clustering & PCA →
Go big — advanced & GenAI projects
The projects that make a portfolio stand out — and teach you how modern AI products are actually built.
RAG chatbot over your own docs
Answer questions grounded in your notes or a PDF library — retrieve the right chunks, then let an LLM write the answer.
Builds on: RAG →Fine-tune a small LLM with LoRA
Teach an open model a new task or style cheaply, then evaluate whether it actually improved.
Builds on: Fine-tuning →An AI agent with tools
Build an agent that plans and calls two tools to answer questions it couldn’t on its own.
Builds on: AI Agents →Chatbot with long-term memory
Give an assistant memory across sessions so it remembers you — the difference between a demo and a product.
Builds on: Prompt Engineering →Semantic image search (CLIP)
Search a photo library with plain-English queries by embedding images and text into the same space.
Builds on: Embeddings →Reproduce nanoGPT
Build and train a tiny GPT from scratch, following Karpathy — the single best way to truly get transformers.
Builds on: Transformers →Deploy a model as an API
Wrap a model in FastAPI, Dockerize it, and put it live — a working demo beats a dozen certificates.
Builds on: Deploying a Model →Build an eval suite for an LLM app
Write tests that catch quality regressions before users do — the skill that separates hobby from production.
Builds on: Evals & Guardrails →
Capstone — portfolio centerpieces
One big, finished, public project that proves you can take AI from idea to something people use. Pick one and go deep.
A full-stack AI product
Frontend + API + model, deployed and usable by a stranger. The single most convincing thing on a résumé.
Builds on: Deploying a Model →Enter a Kaggle competition
Pick an active competition, iterate to a real leaderboard score, and write up what worked.
Builds on: First ML Project →Reproduce a research paper
Read a paper and rebuild its core result. Nothing teaches depth like making someone else’s claim run.
Builds on: Transformers →Ship an end-to-end MLOps pipeline
Data → training → tracking → deployment → monitoring, automated and reproducible.
Builds on: MLOps Basics →Contribute to an open-source AI project
Fix a real bug or add a feature to a library you use. Proof you can work in a real codebase.
Builds on: Portfolio & Career →
Built something? Add it to your portfolio, then head toWhat’s next → to keep going.