The Daily Frustration

Let’s face it: typing on mobile screens is clumsy. Virtual keyboards are cramped, autocorrect often changes words we actually meant to type, and typing long sentences on the go slows us down.

What We Built

We built SmartTyping, a lightweight, predictive text engine designed to guess your next word before your finger even leaves the screen.

• Type Less, Say More: Instead of spelling out every single word, users can just tap our suggestions to speed through sentences.
• Fewer Typos: By shifting the typing experience from spelling things out to picking options, we naturally cut down on errors.
• Built for Integration: It’s designed to drop right into mobile messaging apps, customer chat boxes, or search bars.

Gathering the Data

To teach our app how people actually talk, we fed it a massive dataset of real-world text from blogs, news outlets, and millions of tweets.

Streamlining the Model

• Smart Sampling: We used a neat 5% slice of the overall dataset. This gave us all the common language patterns we needed without bloating the application.
• Deep Cleaning: We lowered all text, stripped out URLs, punctuation, and random characters so the engine only focuses on actual words.
• Building the Dictionary: We grouped the words into chains of text—looking at single words, word pairs (Bigrams), groups of three (Trigrams), and blocks of four (Quadgrams).

To keep things lightning-fast, the engine uses a logic called Stupid Back-off. It sounds funny, but it’s an incredibly efficient way to guess context in milliseconds:

1. The Core Guess (Quadgrams): The engine looks at the last 3 words you typed. If it recognizes the pattern, it gives you the most statistically likely 4th word.
2. Step-Down Logic (Trigrams): If it doesn’t recognize those 3 words, it seamlessly “backs off” to look at just your last 2 words.
3. The Quick Backup (Bigrams): If that fails, it drops down to check just the single last word you typed.
4. The Safety Net (Unigrams): If you type a word the app has never seen before, it simply suggests the most common words in the English language (like “the”, “and”, “is”).

Because of this quick step-down fallback system, the app never gets stuck or gives you an empty blank box.

We didn’t just write a script; we built a clean, working web application that anyone can use right now.

• Instant Updates: The moment you hit the spacebar, the engine recalculates behind the scenes and throws up predictions instantly.
• Transparent Options: It doesn’t just show one word; it displays a neat table of alternative guesses, showing you exactly how confident the model is.
• Zero Lag: We packed and compressed our entire dictionary into a tiny .RData file. This means the server can search millions of rows in milliseconds without skipping a beat.

Test it yourself:

• Live App URL: [https://7o7fo3-kartik-samvedi.shinyapps.io/capstone_project/]

1. Light on Memory, Fast on Speed

You don’t need an expensive graphics card or heavy cloud servers to run this. The entire engine fits inside a tiny file footprint, making it perfect for budget phones or simple web setups.

2. Keeps the Real Context

By looking up to 4 words back, our engine understands conversational phrases, idioms, and common sayings. It naturally recommends words that feel right in context.

3. What’s Next?

• Personalized Typer: Teaching the app to adapt to a user’s specific slang, emojis, and vocabulary over time.
• Fat-Finger Correction: Combining next-word guesses with nearby keyboard layout tracking to fix mistyped letters on the fly.