• Modern mobile interfaces demand intelligent, real-time text completion to reduce typing effort and cognitive load.
• Goal: Develop an accurate, ultra-low latency predictive text engine capable of running seamlessly on web and mobile environments.
• Value Proposition:
  • Sub-millisecond response times.
  • Zero heavy server dependencies.
  • Scalable statistical Natural Language Processing (NLP) architecture.

• Training Corpus: English text sourced from SwiftKey (en_US.blogs.txt, en_US.news.txt, en_US.twitter.txt) totaling > 4 million lines.
• Cleaning Strategy:
  • Tokenization, lower-casing, URL removal, and punctuation stripping.
  • Removal of profanity and low-frequency sparse tokens.
• Corpus Characteristics:

The predictive engine employs an \(N\)-gram language modeling approach with back-off fallback:

1. Trigram Match (\(N=3\)): Given the last two typed words \((w_{i-2}, w_{i-1})\), search for the candidate \(w_i\) maximizing \(P(w_i \mid w_{i-2}, w_{i-1})\).
2. Bigram Fallback (\(N=2\)): If no matching trigram exists, back off to \((w_{i-1})\) to find the most probable \(w_i\).
3. Unigram Fallback (\(N=1\)): If an unseen term is entered, default to the corpus mode (e.g., “the”).
4. Pruning: Frequency matrices were pruned by removing singleton occurrences, reducing RAM usage by > 85%.

• Features: Clean UI, instant reactive prediction, and algorithmic transparency explaining which \(N\)-gram rule resolved the query.

• Demonstrated Strengths: High out-of-sample accuracy on common phrases with negligible latency.
• Future Enhancements:
  • Semantic context weighting via word embeddings (Word2Vec / GloVe).
  • Adaptive user vocabulary personalization based on local browser cache.
• Live Links:
  • Shiny App: https://jramirezgen.shinyapps.io/next_word_predictor/
  • GitHub Code: https://github.com/jramirezgen/