K(no)wBall | Devpost

Overview
Extension
Extension

Inspiration

We don’t watch that many sports, so whenever big games are happening and our group chats blow up, we usually have no idea what’s going on... and are forced to stay silent. Between the Knicks, the U.S. Open, the World Cup and then also MLB, keeping up felt like a full-time job requiring a decade's worth of hyper-niche player lore, past injury histories, and trade drama.

Googling in the middle of a fast-moving conversation is way too slow: by the time you open another app, skim three articles, and piece together the backstory, the chat has moved on. We wanted to build something that lets us convincingly pretend we know ball without spending 40 hours a week watching match highlights. K(no)wBall was born to take the real-time discourse and hot takes already circulating online and turn them into natural, instant banter right where your chat lives.

What it does

K(no)wBall is a Chrome extension that acts as your real-time sports ghostwriter directly inside your active chat window: helping you know ball when you got no ball.

Instead of reciting boring Wikipedia stats or box scores, K(no)wBall reads the context of your chat, taps into live internet discourse, and generates short, punchy responses tailored to casual text speak across multiple vibes (Safe, Funny, Spicy). With a single click on your preferred response, the extension automatically injects the text straight into your message input field, ready to send—meaning zero app-switching, zero research, and no more awkward silence.

How we built it

Frontend Extension: Built as a lightweight Chrome extension that injects directly into the web chat interface. It leverages DOM Mutation Observers to detect incoming chat nodes asynchronously, extracting recent message context while sanitizing and stripping out personal chat metadata.
Real-Time Data Ingestion: To capture live fan sentiment without paying for expensive official APIs, we used Browserbase to orchestrate headless browser instances. By configuring persistent Context IDs, we preserved our authenticated session state and cookies across runs, enabling our scraper to navigate live X (Twitter) search feeds and pull current discourse without hitting login walls.
LLM & Tone Alignment: We passed the scraped live discourse and chat context into a fast LLM pipeline with carefully tuned few-shot prompts. The system maps the sentiment into conversational text speak—stripping out robotic AI punctuation and corporate phrasing to produce authentic fan banter.
One-Click DOM Injection: Designed a responsive content overlay that allows users to preview suggested lines and inject them directly into the active input element via simulated DOM events.

Challenges we ran into

X (Twitter) Scraping & Auth Barriers: The official X API has no free tier and quickly becomes cost-prohibitive for a hackathon build. However, modern X aggressively blocks unauthenticated scrapers with immediate login barriers. Setting up Browserbase with persistent Context IDs required careful handling of session cookies and state persistence to ensure our scraper could access live discourse consistently without triggering repeated auth challenges.
Low Latency for Fast Group Chats: Group chats move in seconds—a joke or hot take that arrives 15 seconds late is dead on arrival. Bridging headless browser scraping, payload extraction, and LLM generation within a conversational timeframe required aggressive optimization of our scraping queries and LLM parameters.
The "Vibe" Problem: Default LLMs sound like encyclopedias or corporate assistants. Getting the model to produce brief, cynical, or comedic fan banter that matches casual lowercase text speak took extensive prompt engineering and fine-tuning.

Accomplishments that we're proud of

Sub-Second Pipeline & Latency Optimization: Successfully cutting down the turnaround time from reading incoming messages to scraping live discourse and generating ready-to-send responses.
Bypassing the Auth Wall with Browserbase: Getting reliable, authenticated scraping operational on X without relying on expensive enterprise API tiers.
End-to-End Interaction Flow: Progressing iteratively from hardcoded button tests to parsing dynamic values, to automatic one-click input field population, and ultimately automating ambient DOM context reading so the extension tone-matches the conversation automatically.

What we learned

How to manage stateful, authenticated headless browsing at scale using Browserbase sessions and Context IDs.
Techniques for building non-intrusive Chrome extension content scripts that monitor dynamic DOM mutations without impacting host page performance.
The nuances of few-shot prompt design for social mimicry—specifically how to steer LLMs away from formal sentence structures toward believable, internet-native text banter.

What's next for KnowBall

Multi-Platform Support: Expanding native integration beyond browser-based chats to Discord desktop clients, Slack, and WhatsApp Web.
Beyond Sports: Extending the discourse engine to other fast-moving pop-culture niches with high barriers to entry, such as reality TV drama, awards season, and trending tech discourse.
Predictive Pre-fetching: Caching live sentiment for trending games and matches using local fast storage (like Redis) so suggested responses generate virtually instantaneously as soon as a key play occurs.