Typing on mobile devices can be slow and prone to errors. SwiftPredict solves this by providing real-time word suggestions.

• Value Proposition: Increases typing efficiency by over 30%.
• Target Audience: Mobile users and accessibility software developers.
• Goal: Deliver a lightweight, sub-second response prediction engine.

The model was built using the HC Corpora dataset, comprising millions of lines from Twitter, Blogs, and News.

• Sampling: 1% of total data used to optimize Shiny memory limits.
• Cleaning: Used stringi and tidytext to remove numbers, punctuation, and profanity.
• Tokenization: Created N-gram frequency tables (Bigrams, Trigrams, and Quadgrams).

We implemented a Stupid Back-off algorithm, which is the industry standard for speed and efficiency in web apps.

• Step 1: Search Quadgrams for the last 3 words entered.
• Step 2: If no match, “back-off” to Trigrams for the last 2 words.
• Step 3: If no match, check Bigrams for the last 1 word.
• Step 4: Fallback to the most frequent word (“the”) if no match is found.

The application is hosted on shinyapps.io and features a minimalist design.

• Input: Reactive text box that captures user phrases.
• Output: Instantaneous display of the predicted next word.
• Optimization: Data is stored in compressed .rds files, ensuring the app loads in under 2 seconds.

[Insert App Screenshot Here]

SwiftPredict demonstrates that complex NLP models can be deployed as lightweight web applications.

• Live App: [Link to your shinyapps.io]
• Source Code: [Link to your GitHub]
• Next Steps: Integrating “Katz Back-off” for higher accuracy and expanding the dictionary.

Thank you!