Mindshield

Inspiration

Online harassment is a low-frequency but high-impact risk for content creators or influencers.

Most creators will never face large-scale attacks—but when it happens, the psychological damage can be overwhelming and immediate; Also, the imagination of going through it exists day by day, often fills them with fear and anxiety, even affecting their creativity and expression.

It‘s torture that you don’t know whether or when it would happen. So, I thought—what if we could help creators the way insurance helps with other risks? By creating a shared compute pool where users contribute, those who face attacks can make full use of it for protection.

This way, the cost is quite low, and creators gain the mental peace they deserve, knowing that even in a crisis, They have a companion to stand by their side in the fight.

What it does

Mindshield is a browser-based protection extension for anyone who displays themselves on social media, especially for content creators or influencers.

The core of the product helps users deal with negative comments and direct messages on social media platforms. It features two modes: Daily Mode and Crisis Mode. In Daily Mode, it continuously filters comments and direct messages in a low-power way. When abnormal attack patterns are detected, Mindshield automatically switches to Crisis Mode, deploying top-tier models and maximum computing resources to:

Filter different types of negative content in real time
Block malicious or coordinated accounts
provide AI analysis to help response
Preserve evidence for later review or reporting

Crisis Mode stays active until the incident is over.

How we built it

We built Mindshield as a browser extension + cloud AI system.

A lightweight browser extension injects real-time protection into social media pages，which is perceived as floating widget.
Background monitoring runs in low-power mode to minimize cost
When risk thresholds are triggered, the system escalates to high-intensity inference using Gemini models
A cloud dashboard visualizes intercepted content, attack status, and evidence logs

Our comment filtering system integrates Gemini 3 as the deep analysis layer in a three-tier architecture: Rule Engine → ML Classifier (Detoxify) → LLM Analyzer (Gemini).

How Gemini 3 is Used:

Context-Aware Analysis: When rule-based and ML layers produce uncertain results (confidence < 80%), Gemini 3 performs deep semantic analysis to understand nuanced expressions, sarcasm, and context-dependent toxicity that traditional models miss.

Intent Recognition: Gemini analyzes user intent behind comments - distinguishing genuine praise like "I hate how talented you are" (positive) from actual hate speech, reducing false positives significantly.

Multi-language Support: Leveraging Gemini's multilingual capabilities to handle mixed-language comments (Chinese/English) and internet slang that rule-based systems cannot cover.

Structured Output: Using Gemini's JSON mode to return standardized classification results including category, confidence score, severity level, and reasoning - enabling transparent and auditable moderation decisions.

Fallback Intelligence: When upstream layers conflict, Gemini serves as the final arbiter, providing human-like judgment for edge cases.

Key Gemini 3 Features Utilized: Advanced reasoning for context understanding Multilingual text comprehension Structured JSON output generation Low-latency inference for real-time filtering

Through conversation, Gemini 3 was used to complete the full MVP feature list and high-fidelity UI interactions. Because the product requires user login and social media authorization, a demo version was built for the hackathon: the backend and algorithm components are real, while the social media pages are mocked. User authorization and purchase flows are not shown.

Roles & Responsibilities:

Nantsy H: Product concept, PM, and UI
Jingyao Chen: Development — frontend, backend, and database
Yunming Tang: Algorithms — Gemini model invocation logic

Challenges we ran into

Defining “crisis”: Distinguishing between normal negativity and coordinated attacks without false positives.
Psychological UX: Designing feedback that reassures users without drawing attention to harmful content.
Platform differences: Different social platforms allow very different levels of automation and intervention.
User data security and privacy: Given the sensitive nature of personal messages and comments, we must ensure that no data is exposed or misused in the product.
Cultural and linguistic variation: Identifying negative content accurately across different languages and cultural contexts proved challenging.

Accomplishments that we're proud of

Compared to a subscription model, this service mechanism aligns more closely with users' needs
The daily mode features a highly cost-effective filtering mechanism, ensuring affordability and sustainable operations.
Turning an abstract emotional problem into a clear, testable product system

What we learned

The product is somewhat dependent on the policies of the social media platforms themselves. It requires a comprehensive and well-balanced combination of feasible approaches: web scraping, pure visual solutions, and API integration.
Escalation-based AI systems can significantly reduce cost without sacrificing protection.
In daily mode, the product's functionality is not easily apparent. It's important to consider strengthening user trust through UI and other means.