Gewee

Inspiration

The modern web is a battleground of information overload and manipulative "dark patterns". Most AI assistants live in a sidebar, forced to wait for user input, but I envisioned a proactive partner that understands the "mission" of a webpage before the user even asks. Inspired by the need for digital clarity, I built Gewee to act as a cognitive layer that filters noise, automates tedious drafting, and enhances human intent directly within the browser's operational framework.

What it does

Gewee is a mission-oriented browser agent that autonomously adapts its persona—Tutor, Muse, or Guardian—based on real-time site context.

Tutor Mode: Active on technical documentation (e.g., Wikipedia, Hugging Face), it highlights complex terms and provides instant educational layers.

Muse Mode: Active on creative and social platforms (e.g., LinkedIn, YouTube), it drafts contextually appropriate social responses directly in comment fields.

Guardian Mode: A security-first persona that activates on login pages (e.g., GitHub) to signal active data masking and protection against "dark patterns".

Interactive Suite: Beyond personas, it offers "Deep Research," "Fact Check," and "Visualize" tools to process multimodal information like text and code simultaneously.

How we built it

The extension is built on Chrome Manifest V3 using JavaScript (ES6+) and Gemini 3 Flash for high-speed, autonomous reasoning. We utilized Shadow DOM for UI isolation to ensure Gewee’s interface never clashes with a website's native styles. The agentic logic uses real-time DOM snapshots and Gemini 3’s Thinking Levels to triage tasks—using "High" thinking for technical explainers and "Minimal" for rapid response drafting.

Challenges we ran into

One major hurdle was Spatial Intelligence—ensuring that injected buttons didn't block critical website labels, which we solved with a custom collision-detection algorithm. We also fought Layout Thrashing; initially, extracting page data caused stutters, so we refactored to a Zero-Reflow system using textContent and requestAnimationFrame to maintain 60fps performance even on heavy documentation. Finally, balancing persona triggers required a strict Priority Override system to ensure security (Guardian) always took precedence over creative suggestions on sensitive pages.

Accomplishments that we're proud of

I am incredibly proud of achieving a 40% reduction in token waste through our "Standby Mode" and character-limited thought bubbles, proving that AI agents can be powerful without being expensive or noisy. Additionally, the seamless integration of Human-in-the-Loop design—where the agent only "wakes up" with high-value insights when a knowledge gap is detected—sets it apart from typical, always-on chatbots.

What we learned

Building Gewee taught me the importance of Surgical AI interaction. We learned that highlighting a single technical term with an "Explain This" button is often more valuable than a full-page summary. I also gained deep experience in Chrome Storage security, moving from hardcoded configurations to a user-controlled Options UI that keeps API keys secure and decentralized.

What's next for ByJir

The vision for Gewee (ByJir) is to expand the persona library to include a "Researcher" mode for cross-tab synthesis and a "Developer" mode that integrates directly with local IDEs. I plan to further leverage Gemini 3’s Multimodal Vision to detect accessibility issues on websites in real-time, making the web not just smarter, but more inclusive for everyone.

Built With

• chrome-extension-api-(manifest-v3)
• chrome.storage
• gemini-3-flash-api
• html5/css3
• javascript-(es6+)
• local
• shadow-dom