Type faster. Keep the user in control.

• Accepts any multi-word English phrase
• Lets the user request 1, 2, or 3 suggestions
• Ranks each candidate and displays an estimated probability
• Works in a browser through a responsive Shiny interface

Product value: one prediction is an answer; three ranked predictions are a useful choice.

1. Normalize and tokenize the phrase
2. Match the last two words against trigrams
3. Back off to bigrams when context is sparse
4. Fill remaining positions using unigram frequency
5. Weight, rank, deduplicate, and normalize candidate scores

Probability: conditional frequency × backoff weight ? normalized displayed probability.

• 3 backoff levels and 1, 2, or 3 selectable predictions
• Precomputed n-gram tables keep prediction lightweight
• Compact model data supports fast loading and deployment
• Approximately 6 ms per prediction in a local 1,000-call benchmark
• Graceful fallback ensures a result for unfamiliar input

The architecture can replace the prototype corpus with a production-scale model.

Choose the number of suggestions, enter a phrase, and press Predict next words.

Launch the live Shiny app

The interface returns ranked cards containing the word and its percentage.

Next iteration: larger training data, held-out evaluation, interpolated probabilities, slang handling, and personalization.