An AI that plays Hangman by predicting the next letter using a custom-trained Transformer model and real-time gameplay in a retro-style Streamlit app.

Try it live 👉 Play Here!

This project started as a take-home exercise for an interview process and evolved into a full-featured, character-level AI game. The AI plays classic Hangman using a Transformer model trained on masked word samples and gameplay context with no handcrafted logic or heuristics, just pure model-driven inference.

Training set: ~84,000 words (SCOWL 70)

Test set: ~8,000 words (SCOWL 60)

Gameplay: Based on a filtered list of ~3,000 hand-cleaned words from the test set

Win rate: Achieves 58% accuracy on the full test set (completely unseen during training)

A PyTorch Transformer encoder trained to predict the next letter in masked Hangman-style words.

19 auxiliary features (e.g., mask structure, guessed letters, remaining guesses)

4 multi-hot feature vectors (vowels, consonants, guesses, misses)

Masked entropy score for uncertainty awareness

PyTorch-based training loop

Early stopping, validation tracking

Regularisation: token dropout, label smoothing, weight decay

Trained on GPU

Built with Streamlit, the web app provides:

Real-time interactive play

Local model inference (no backend required)

Retro UI, responsive layout (best played on desktop, mobile works in landscape)

I'm actively exploring new features and would love suggestions from the community! Here are some ideas already on the roadmap:

Whole-word guesses: Let users attempt full-word guesses (like real Hangman), with a penalty mechanism that disadvantages the AI if guessed correctly

Difficulty settings: Adjustable game difficulty by altering the word list (e.g. rare vs common words)

AI explainability: Show the AI’s top candidate letters or confidence scores (optional toggle for the curious)

Human vs AI stats: Track win/loss records across sessions or leaderboards

Word definitions: Option to reveal the word’s meaning after the game (educational bonus!)

Mobile UX improvements: Full-screen mode and better gesture handling on phones/tablets

Alternative models: Experiment with distilled models or LSTMs for faster loading/inference

👂 Got an idea? Open an issue or submit a PR! I'm keen to hear what features you'd enjoy or how to make the AI smarter/fairer/more fun.