This repository contains my completed project work for Harvard CS50's Introduction to Artificial Intelligence with Python.
I completed all 12 CS50AI projects, covering the full course journey from classical search algorithms to modern transformer-based natural language processing. Completing these projects earned me the CS50AI Certificate from Harvard University, validating my hands-on understanding of core artificial intelligence concepts and my ability to implement them in Python.
This repository is more than a course archive.
It is a portfolio of AI foundations.
Each project represents a different layer of intelligent systems: how machines search, reason, predict, learn, and understand language.
CS50AI covers the core ideas behind modern artificial intelligence, including:
- Search algorithms
- Game-playing AI
- Knowledge representation
- Logical inference
- Uncertainty
- Bayesian networks
- Markov models
- Constraint satisfaction problems
- Supervised machine learning
- Reinforcement learning
- Neural networks
- Natural language processing
- Transformer-based attention models
Completing all 12 CS50AI projects earned me the CS50AI Certificate from Harvard University.
This certificate represents my completion of Harvard's artificial intelligence curriculum, including hands-on work across search, logic, probability, optimization, machine learning, neural networks, and natural language processing.
Through these 12 projects, I strengthened my understanding of how intelligent systems think, search, reason, predict, and learn.
This repository represents my foundation in:
-
Classical AI
- Search
- Logic
- Constraint satisfaction
- Knowledge-based agents
-
Probabilistic AI
- Bayesian inference
- Markov chains
- Probability distributions
- Uncertainty modeling
-
Machine Learning
- Classification
- Model evaluation
- Reinforcement learning
- Neural networks
-
Natural Language Processing
- Context-free grammar parsing
- Masked language modeling
- Transformer attention analysis
This is where I connected AI theory with real implementation.
Not just “what is AI?”
But how does AI actually work under the hood?
Concept: Search Algorithms
Core Technique: Breadth-First Search
This project finds the shortest connection between two actors using movie co-star relationships.
The program models actors as states and movies as actions. Using breadth-first search, it determines the minimum number of “degrees of separation” between two actors.
- Graph search
- Shortest path discovery
- Breadth-first search
- State-space modeling
- Frontier-based exploration
Concept: Game-Playing AI
Core Technique: Minimax
This project implements an unbeatable Tic-Tac-Toe AI using the Minimax algorithm.
The AI evaluates all possible future board states and selects the optimal move for the current player. With correct play, the AI cannot be beaten.
- Adversarial search
- Minimax decision-making
- Terminal state evaluation
- Utility functions
- Optimal game strategy
Concept: Knowledge Representation
Core Technique: Propositional Logic + Model Checking
This project solves Knights and Knaves logic puzzles.
Each character is either a knight who always tells the truth or a knave who always lies. The AI represents puzzle statements using propositional logic and uses model checking to determine which conclusions must be true.
- Logical symbols
- Knowledge bases
- Entailment
- Model checking
- Propositional logic
Concept: Knowledge-Based AI
Core Technique: Logical Inference
This project builds an AI agent that plays Minesweeper by reasoning about safe cells and mines.
The AI stores logical sentences about the board and uses inference rules to mark cells as safe or dangerous. When no guaranteed safe move exists, it falls back to a random move.
- Knowledge-based agents
- Inference from constraints
- Logical sentence simplification
- Safe move selection
- Uncertainty handling
Concept: Probabilistic Models
Core Technique: Markov Chains + Iterative Ranking
This project implements the PageRank algorithm used to rank web pages by importance.
The program calculates page importance using two approaches:
- Sampling from a random surfer model
- Iteratively applying the PageRank formula
PageRank models the web as a graph and estimates how likely a random user is to land on each page.
- Markov chains
- Random surfer model
- Damping factor
- Iterative convergence
- Link analysis
- Probability distributions
Concept: Uncertainty
Core Technique: Bayesian Inference
This project estimates the probability that individuals carry a genetic trait based on family relationships and observable evidence.
The AI uses joint probability, inheritance rules, mutation probability, and normalization to calculate probability distributions for each person's gene copies and trait expression.
- Bayesian networks
- Joint probability
- Conditional probability
- Genetic inheritance modeling
- Normalization
- Probabilistic inference
Concept: Constraint Satisfaction Problems
Core Technique: Backtracking Search + Arc Consistency
This project generates crossword puzzles by assigning words to crossword slots while satisfying constraints.
The AI treats each word slot as a variable, each possible word as a domain value, and overlaps between words as binary constraints.
- Constraint satisfaction
- Node consistency
- Arc consistency
- AC-3 algorithm
- Backtracking search
- Minimum remaining values heuristic
- Least-constraining value heuristic
Concept: Supervised Machine Learning
Core Technique: k-Nearest Neighbors Classification
This project predicts whether an online shopping user will complete a purchase.
The model uses user session data such as page visits, durations, bounce rates, traffic type, visitor type, weekend activity, and other features to classify whether revenue will be generated.
- Classification
- k-nearest neighbors
- Feature preprocessing
- Train-test split
- Sensitivity
- Specificity
- Model evaluation
Concept: Reinforcement Learning
Core Technique: Q-Learning
This project builds an AI that teaches itself to play Nim through self-play.
The AI learns Q-values for state-action pairs. Over thousands of training games, it discovers which moves lead to winning outcomes and which moves should be avoided.
- Reinforcement learning
- Q-learning
- Rewards and penalties
- Exploration vs exploitation
- Epsilon-greedy strategy
- Self-play training
Concept: Computer Vision
Core Technique: Convolutional Neural Networks
This project uses TensorFlow to classify traffic signs from images.
The AI trains on the German Traffic Sign Recognition Benchmark dataset and learns to identify 43 categories of road signs using image preprocessing and a neural network model.
- Deep learning
- Computer vision
- Image classification
- Convolutional neural networks
- TensorFlow
- OpenCV
- Model experimentation
Concept: Natural Language Processing
Core Technique: Context-Free Grammar Parsing
This project parses English sentences and extracts noun phrase chunks.
The AI uses context-free grammar rules to generate parse trees and identify meaningful noun phrases inside a sentence.
- Natural language processing
- Tokenization
- Context-free grammar
- Parse trees
- Noun phrase extraction
- NLTK
Concept: Transformer-Based NLP
Core Technique: BERT Masked Language Modeling + Attention Visualization
This project uses BERT to predict masked words in a sentence.
It also generates visualizations for BERT’s self-attention heads, helping analyze how different tokens attend to each other inside a transformer model.
- Transformers
- BERT
- Masked language modeling
- Self-attention
- Attention head visualization
- Hugging Face Transformers
- NLP interpretability
After completing the full CS50AI project set, I chose three projects to take beyond the course specification.
These were not just submitted as academic assignments.
I selected them because they represent three different pillars of artificial intelligence:
- Reinforcement Learning
- Graph-Based Ranking Systems
- Transformer-Based Natural Language Processing
The goal was simple:
Take strong CS50AI foundations and rebuild selected projects into more polished, industry-standard, production-ready portfolio projects.
These enhanced versions combine what I learned from:
- Harvard CS50AI
- My B.Tech Computer Science and Engineering degree
- Data Structures and Algorithms
- Machine Learning coursework
- Deep Learning coursework
- Software Engineering principles
- Cloud and DevOps project experience
- Previous full-stack and AI portfolio projects
This is where the course work became something bigger.
Not just assignments.
Portfolio-grade AI systems.
The original CS50AI Nim project focuses on Q-learning through self-play.
For the enhanced version, I expanded the idea into a more complete reinforcement learning project that better demonstrates how agents learn from experience.
The course version implements:
- Nim game environment
- Q-value storage
- Q-learning update rule
- Epsilon-greedy action selection
- Self-play training loop
- Human vs AI gameplay
The enhanced version can showcase:
- Cleaner RL environment structure
- Training analytics
- Win-rate tracking
- Q-table inspection
- Configurable training episodes
- Strategy evaluation
- Improved CLI or UI experience
- Documentation explaining reinforcement learning in simple terms
Nim is a perfect foundation for understanding how machines learn through trial and error.
It shows the heart of reinforcement learning:
An agent takes actions, receives feedback, and slowly learns better decisions.
That idea powers robotics, game AI, recommendation systems, autonomous systems, and real-world decision engines.
The original CS50AI PageRank project implements Google’s classic ranking algorithm using both sampling and iterative convergence.
For the enhanced version, I treated PageRank as more than a search engine concept.
I rebuilt it as a graph intelligence project.
The course version implements:
- Corpus crawling
- Link graph construction
- Transition model
- Random surfer sampling
- Iterative PageRank calculation
- Damping factor handling
- Convergence thresholding
The enhanced version can showcase:
- Graph visualization
- Interactive ranking exploration
- Custom webpage or document corpus support
- Comparison between sampling and iteration
- Ranking explainability
- Network analysis metrics
- Cleaner modular architecture
- Potential dashboard or API layer
PageRank is one of the most important algorithms in Internet history.
But the deeper lesson is bigger than web pages.
PageRank teaches how importance flows through a network.
That applies to:
- Search engines
- Citation networks
- Social graphs
- Recommendation systems
- Knowledge graphs
- Dependency graphs
- Trust and influence scoring
View Enhanced PageRank Project
The original CS50AI Attention project uses BERT to predict masked words and generate attention diagrams.
For the enhanced version, I focused on making transformer attention more understandable, visual, and useful.
The course version implements:
- BERT masked word prediction
- Mask token detection
- Attention score color mapping
- Attention diagram generation
- Human interpretation of attention heads
- Analysis of token relationships
The enhanced version can showcase:
- Cleaner attention visualizations
- Better explanation of transformer internals
- Layer/head comparison
- Token-to-token attention exploration
- Example sentence library
- Interpretability notes
- Streamlit or web-based interface
- Stronger NLP portfolio documentation
Modern AI is powered by transformers.
Large language models, coding assistants, search systems, chatbots, summarizers, and AI agents all rely on attention mechanisms.
This project helped me understand a key question:
When a model reads a sentence, what is it paying attention to?
That question is central to NLP, explainable AI, and modern LLM systems.
View Enhanced Attention Project
I chose Nim, PageRank, and Attention because together they form a powerful AI learning arc.
Nim represents agents that improve through feedback.
This connects to:
- Reinforcement learning
- Decision-making systems
- Autonomous agents
- Game AI
- Optimization through self-play
PageRank represents intelligence over graphs and relationships.
This connects to:
- Search engines
- Recommendation systems
- Graph analytics
- Network science
- Authority and trust scoring
Attention represents modern NLP and transformer-based intelligence.
This connects to:
- Large language models
- Chatbots
- Semantic search
- Prompt engineering
- AI interpretability
- Context-aware systems
Together, these three projects show a strong range:
| Project | AI Area | Main Skill Demonstrated |
|---|---|---|
| Nim | Reinforcement Learning | Learning optimal actions through experience |
| PageRank | Graph Algorithms | Ranking importance across connected systems |
| Attention | NLP / Transformers | Understanding language through contextual attention |
This combination shows that I am not only learning AI from one angle.
I am building across the full stack of AI thinking:
decision-making, ranking, reasoning, and language understanding.
- Python 3.12
- Search
- Minimax
- Propositional logic
- Model checking
- Bayesian inference
- Markov chains
- Constraint satisfaction
- k-nearest neighbors
- Q-learning
- Neural networks
- Context-free grammar
- Transformer attention
- pygame
- scikit-learn
- TensorFlow
- OpenCV
- NLTK
- Hugging Face Transformers
- NumPy
- pandas
- Pillow
CS50AI/
│
├── degrees/
│ └── Shortest-path actor connection search
│
├── tictactoe/
│ └── Minimax Tic-Tac-Toe AI
│
├── knights/
│ └── Logic puzzle solver
│
├── minesweeper/
│ └── Knowledge-based Minesweeper AI
│
├── pagerank/
│ └── PageRank sampling and iteration
│
├── heredity/
│ └── Bayesian genetic inheritance inference
│
├── crossword/
│ └── Constraint satisfaction crossword generator
│
├── shopping/
│ └── k-nearest neighbors purchase prediction
│
├── nim/
│ └── Q-learning Nim AI
│
├── traffic/
│ └── CNN traffic sign classifier
│
├── parser/
│ └── Context-free grammar parser
│
├── attention/
│ └── BERT masked language model attention visualizer
│
└── README.md
- Degrees
- Tic-Tac-Toe
- Knights
- Minesweeper
- PageRank
- Heredity
- Crossword
- Shopping
- Nim
- Traffic
- Parser
- Attention
This repository demonstrates my ability to:
- Implement AI algorithms from scratch
- Translate mathematical concepts into working code
- Build intelligent agents
- Work with probability and uncertainty
- Design search-based solutions
- Train and evaluate machine learning models
- Build neural networks for image classification
- Use NLP tools for parsing and masked language modeling
- Analyze transformer attention behavior
- Explain technical concepts clearly through documentation
CS50AI helped me understand that artificial intelligence is not magic.
It is built from layers of ideas.
Search.
Logic.
Probability.
Learning.
Language.
Each project added one more piece to the puzzle.
And after completing the course, I wanted to go one step further.
That is why I chose Nim, PageRank, and Attention as my enhanced projects.
Because they represent where AI becomes real:
- AI that learns
- AI that ranks
- AI that understands language
That is the bridge between coursework and industry.
| Enhanced Project | Focus Area | Repository |
|---|---|---|
| Nim | Reinforcement Learning | View Repo |
| PageRank | Graph Ranking / Search Intelligence | View Repo |
| Attention | Transformer NLP / Interpretability | View Repo |
This repository is based on my learning from Harvard CS50's Introduction to Artificial Intelligence with Python.
Some distribution files, datasets, and starter code belong to CS50AI. My work focuses on the required implementations, algorithmic logic, experimentation, and project documentation.
Large datasets such as the traffic sign image dataset are not included directly in this repository when restricted by assignment or file-size guidelines.
Mitra Boga
This repository is for educational and portfolio purposes.
Please respect Harvard CS50's academic honesty policy and use this repository as a learning reference, not as a direct submission source.