Team Cornerstone: AI-Enhanced Creator Support

Project Inspiration

Our team, Team Cornerstone, is a mix of avid TikTok users and those less familiar with the platform. Through discussions, we realized a common frustration: the overwhelming presence of low-quality or “content-grab” videos. This inspired us to create a system that incentivizes creators to improve content quality while providing actionable feedback for growth.

What It Does

Our app functions as an add-on for TikTok creators. After a video is uploaded, our AI evaluates it based on multiple criteria:

Clarity
Educational Value
Delivery
Audio and Visual Quality
Originality
Video Length
Compliance with Guidelines

Creators receive a Content Quality Score alongside actionable insights, encouraging the production of higher-quality, engaging content.

How We Built It

We leveraged the Lynx framework to integrate our app seamlessly with the TikTok workflow.

ML Model: Trained to assess videos using engagement metrics, content features, and compliance checks.
Data Processing: Implemented with Python, LightGBM, and scikit-learn.
Integration Challenges:
- Learning Lynx, which was new to the team.
- Handling video uploads creatively since <input> fields are not natively supported.

Key Features

Multi-Stream Reward Distribution

Creators are rewarded based on:

Creator Reward = Consumer Gifts (40%) + Advertisement Revenue (30%) + Creator Fund (30%)

Advertisement revenue considers views and watch time.
Creator Fund is weighted by AI-generated content quality scores.

Advanced AI-Powered Content Quality Scoring

Overall Content Quality Score = Engagement Quality (75%) + Content Quality Score (25%)

Inputs for Scoring:

Likes, Shares, Comments → audience engagement
Watch Time → content retention
Content Length → context for engagement
Creator Reputation → prior compliance and behavior

Preprocessing:

Min-Max Scaling to normalize feature ranges (saved as scaler.pkl).

Model:

LightGBM: Fast, explainable, handles non-linear interactions.
Saved for real-time scoring without frequent retraining.

Real-Time Fraud Prevention

Daily gift limits per user
Immutable audit log of flagged transactions
Detection of suspicious gifting patterns (e.g., repeated self-gifting)
Manual flagging capability

Anomaly Detection Model

Purpose: Detect unusual or fraudulent gifting behavior.

Synthetic Data Generation:

20 users simulated (user1–user20)
Actions include normal gifting, suspicious gifting (800–1500 tokens), and potential gaming (400–700 tokens)
Timestamps jittered ±1 hour for realism

Feature Engineering:

Amount – token value of gift
User Numeric – encoded user ID
Action Code – numeric mapping of action types
Timestamp – epoch time of transaction

Model Architecture:

Isolation Forest (tree-based ensemble for anomaly detection)
Hyperparameters: contamination=0.2, random_state=42
Outputs: 1 → normal, -1 → anomaly

Challenges We Overcame

Mastering the Lynx framework.
Handling video uploads creatively.
Designing models that are both accurate and explainable.

Accomplishments

Built a functional AI-driven content scoring system.
Developed a reward distribution mechanism combining engagement, revenue, and quality.
Successfully simulated and detected anomalous gifting behaviors.

Key Learnings

Effective AI evaluation requires combining multiple signals beyond raw engagement.
Real-time anomaly detection can prevent fraud without hindering user experience.
Integration with existing platforms like TikTok demands both technical creativity and adaptability.

Technologies Used:

Frontend: React, ReactLynx, TypeScript Backend: FastAPI, Python, PostgreSQL, Prisma ORM Machine Learning / AI: TensorFlow, Keras, LightGBM, scikit-learn, OpenCV, NumPy, Pandas APIs & Integrations: Lynx framework for connecting frontend with backend and ML services