Typing is slow. Prediction is fast.

• Mobile keyboards, search boxes, and chat apps all rely on next-word prediction to save users keystrokes.
• Our app demonstrates a lightweight prediction engine that runs entirely in the browser session — no GPU, no cloud LLM, sub-second response.
• Trained on millions of real sentences from Twitter, news, and blogs (the SwiftKey corpus).

Type “happy new” → the model instantly suggests “year”.

We use an n-gram model with Stupid Backoff:

1. Pre-process: lowercase, strip punctuation/numbers/URLs, remove profanity.
2. Build tables: count the most frequent 2-, 3-, and 4-word sequences; keep the top 5 continuations per prefix.
3. Predict (backoff): match the longest possible prefix, falling back to shorter n-grams with a 0.4 discount per level.

\[score = count(prefix \rightarrow word) \times 0.4^{\,levels\ dropped}\]

This balances accuracy (long context wins) with coverage (short context always answers). The final model compresses to a few MB so it loads instantly on shinyapps.io.

Three controls, one result:

• Phrase box — type any word or phrase.
• Slider — choose how many predictions (1–5) you want returned.
• Predict button — returns ranked next-word suggestions as pills.

A Home tab holds the predictor; an About tab explains the method and gives instructions. Clean, responsive, mobile-friendly UI.

Tested on 5 held-out Twitter/news phrases — a prediction returned for every one.

• Fast: pure lookup tables — predictions in milliseconds.
• Small: model ships in megabytes, not gigabytes.
• Accessible: anyone with a browser can use it — no install.
• Extensible: swap in higher-order n-grams or neural models behind the same interface.

Try it live: https://bkamra56.shinyapps.io/nextword/

Source & deck: https://rpubs.com/bkamra56/nextword

Questions welcome.

• App: https://bkamra56.shinyapps.io/nextword/
• Slides: https://rpubs.com/bkamra56/nextword

Built with R, Shiny, and R Markdown for the Data Science Capstone.