Hi everyone. My name is ELA Mo and I'm happy to be presenting here at the JavaScript conference 2025. I wanted to start with a quick introduction about myself. I currently work for the Wells Fargo Bank and I have around 14 years of hands-on experience in advanced data analytics. Credit risk strategy and predictive modeling across financial services, marketing, and customer intelligence. I also have extensive experience in developing machine learning driven fraud detection strategies, segmentation models and campaign optimization frameworks that help improve ROI. Reduce risk exposure and enhance operational efficiency. The topic I'm planning to present today is client side, NLP, JavaScript, text Analytics for user Behavior Intelligence. I'm actually pretty excited today to be talking about how we are using client side natural language processing, NLP, with JavaScript to understand user behavior in real time without sending data to a server. This talk is about rethinking how we do text analytics, moving intelligence closer to the user inside the browser where privacy, latency and interact with all matter most. I'll walk you through why this shift is important, the architecture we used, the tools and techniques involved, and that what kind of insights it unlocks for product teams and developers. Let's jump to the next slide. This slide talks about the hidden data problem. Let's start with the challenge. Traditional analytic tools do a great job of tracking what users do. Clicks, scrolls and time on each page, but not why they behave that way. We are capturing millions of user interactions, but missing the deeper layer of intent, hidden in their language. Every form, submission, chat message, search query, and feedback comment contains behavioral signals. Which are emotion, intent, confusion, satisfaction. For example, a user might click contact support, but the message they write reveals frustration, confusion, or even shown intent. Those insights are hidden inside unstructured text, and historically we have ignored them because crossing language was expensive and required server side pipelines, that's the problem we set out to solve. Moving on to the next slide. So this slide talks about why process NLP client side. So why bring NLP to the client? There are three main drivers, privacy, realtime feedback, and efficiency. We can start talking about privacy in a world of data regulations and user trust concerns. Sending raw text to a server is risky. Client said, NLP means the text never leaves the browser. No data leakage, no compliance headaches. The next one is real time feedback. Processing happens instantly as the user types. We can detect sentiment, hesitation, or frustration and adapt the interference interface in a real time. Maybe offer help highlight CTA or are just tone dynamically. The next one is efficiency. By offloading computations to the user's device, we drastically reduce server costs and network latency. With GPO acceleration in browsers, users essentially bring their own compute so we get privacy, immediacy, and scalability. The Holy Trinity of modern analytics. Moving on to the next slide. This slide talks about the JavaScript NLP toolkit. Now let's talk about the tools that make all this possible. First, natural and compromise. These are lightweight NLP Libraries for tokenization, part of speech tagging and entity recognition that run directly in the browser with zero external dependencies. Second TensorFlows js. This brings the power of pre-trained machine learning models to JavaScript using WebGL for GPU acceleration. We use it for sentiment scoring and text classification. And third web workers and index DB web workers keep heavy NLP computations off the main thread, so the UI never freezes. While Index DB stores results locally for offline access or historical trend analysis, together, these tools let us replicate what used to require a Python backend all within the browser. Moving on to the next slide. This slide gives the architecture overview. Here's a high level view of the architecture. The design goal was to isolate NLP processing from user interaction to keep everything responsive. Web workers handle all the model loading and text analysis in the background. Service workers support offline capabilities, caching models and data, so it still works without connectivity. Index DB serves as local storage for processed analytics, letting us analyze behavior over time, even across sessions. And finally, results are funneled back to a centralized state management layer, think Redux or context API, which updates the react components in real time. This architecture mimics distributed computing, but at the browser level, light modular and event driven. Moving on to the next slide. This slide talks about building realtime sentiment analysis. So how does realtime sentiment analysis actually work? Here's a process. The first step, we load a pre-trained tensor flow dot js model at App startup, usually a lightweight transformer. Or LSTM variant optimized for client use. Next, when a user types into a chat feedback box or a form, we capture the input events, but we use debouncing, so we are not analyzing every keystroke. The next step is the text gets tokenized and vectorized into numerical. Representations and that's passed to the model for in inference, which outputs a sentiment score. Finally, we visualize that score right in the UI may be a color indicator emoji or sentiment garge that updates as they type. This gives immediate emotional feedback and provides continuous sentiment streams that pro product teams can analyze. Moving on to the next slide. This slide talks about user segments hidden in text. Once we have enough text, data patterns start to emerge that reveal distinct user segments. There are different types of user segments hidden in text. First one is hesitant users. They use language full of uncertainty. Words like maybe I think, or questions. They're undecided and need reassurance. The next set of segments are competitive switchers. These users explicitly mention other products or pricing. They're evaluating options and can be won over with targeted offers. The next segment is high intent prospects. They use decisive action oriented language, like ready to buy, need integration, start tomorrow. These are your hottest leads. The next one is at-risk users negative sentiment spikes complaints or cancellation keywords. These users signal churn risk early by categorizing users. This way we can align messaging and support to each behavior type in real time. Moving on to the next slide. This slide talks about topic modeling with JavaScript beyond user sentiment. We also want to understand what people are talking about. That's where topic modeling comes in. We use JavaScript implementations of algorithms like key means or hierarchical clustering to group similar messages without any labels. Each text snippet is vectorized using T-F-I-D-F or word embeddings and dimensionality reduction helps us visualize them in clusters, say billing issues, feature requests. Or performance feedback. What's powerful here is that this runs dynamically as new messages arrive, clusters update automatically. It's like having live market research running inside your product interface. Moving on to the next slide. This slide talks about practical implementation text, pre-processing pipeline. Before any model can make sense of text, we need to clean and structure it. Our pre-processing pipeline looks like this. Normalize, tokenize, vectorize, and analyze. When we talk about normalize, it basically converts everything to lowercase, strip punctuation and trim white space tokenize, split text into words, remove common stop words like the and is, and apply stemming or limitation. A vectorize convert words into numerical vectors using TF IDF, or embeddings. Analyze feed vectors into NLP models for classification or clustering. Each stage is modular and usable, which means developers can swap algorithms or add domain specific steps very easily. Moving on to the next slide. This slide talks about building responsive react components. A big part of making this user friendly is how we display results. We built reusable react components that visualize sentiment and topics in real time. Custom hooks manage background communication with web workers, so react only updates when. Fresh data arrives. We use context providers to share NLP results across the app. So sentiment data in one component can influence recommendations or UI color schemes elsewhere. For vis visualization, we integrated D three Js perfect for generating dynamic sentiment timelines, word clouds, and cluster graphs. The focus is on smooth interactivity, loading states, transitions, and animations that make analytics feel integrated, not intrus. Moving on to the next slide. This slide talks about feature engineering text to metrics. Once we process the text, the next step is to turn linguistic insights into actionable metrics. We do this in four layers, linguistic features, behavioral metrics. Aggregate profiles and actionable scores. Linguistic features, so this is basically sentiment, PO polarity tone, readability, entity density and emotional intensity, behavioral metrics, how often users write, how quickly they respond, and whether their tone changes over time. Aggregate profiles, combining text and behavior into holistic user fingerprints that capture engagement style. Actionable scores things like satisfaction index, churn, probability or intent confidence. This allows us to move from raw language to measurable behavioral in intelligence. Moving on to the next slide. This slide talks about performance optimization strategies. Client side NLP is powerful, but we can be resource intensive. We use several optimization tactics like lazy loading, batch processing, debouncing and caching. We only for lazy loading. We only download heavy tensor flow models when they're actually needed. Keeping the initial bundle, small batch processing, instead of analyzing text one snippet at a time. We queue and process in batches maximizing GPU throughput. Debouncing. We avoid reprocessing partial sentences. Analysis only happens after the user pauses typing. Caching results are stored in indexed DB using hashtag keys. So repeated text isn't reanalyzed. These techniques keep the experience seamless, lightweight on both CPU and memory while still providing instant insights. Moving on to the next slide. This slide talks about data privacy and security concentrations. Privacy isn't an afterthought. It's a design principle. Everything happens locally. User text never leaves the device. Before analysis users give explicit consent through an opt-in prompt that explains exactly what's being analyzed and stored. We allow users to view, export, or delete their stored analytics data at any time. Sensitive text in Index DB is encrypted and inco in incognito or private browsing mode analytics are automatically disabled. This approach shows that advanced AI doesn't have to come at the cost of user trust. We're going to the next slide. This slide talks about real world applications. Let's talk about where this approach adds value in production environments. Some examples are support, chat enhancement. Search intent analysis, feedback classification. For support chart enhancement, we can detect frustration or urgency automatically and route high priority messages to live agents search intent analysis. This classifies search terms to identify what users are truly looking for and where your content gaps are. Feedback, classification, this automatically tags user feedback into categories like. Bugs, features and praise. Saving hours of manual sorting because all this happens at the client side. It's scalable across millions of users without additional backend load, and every user's data remains private. The, this concludes the presentation. Thank you for your time and attention. Clients at NLP represents a big shift from centralized data analysis to edge intelligence where users browsers become mini analytics engines. It's faster. Safer and far more personal. As developers, we now ha have the tools to combine NLP web APIs and frontend frameworks to create experiences that truly understand users. In context, I would like to discuss how you might apply these ideas in your own products, whether it's improving engagement. Personalization or user trust. I would like to thank thank the conference for giving me this opportunity and I hope you all have a good day. Thank you.