FairTik

Inspiration

The creator economy is booming, yet fairness and transparency in revenue sharing remain unsolved problems. Many creators feel undervalued, while users often distrust opaque reward systems. We set out to reimagine this process, and we want to build a solution that makes revenue distribution not only automated and transparent, but also aligned with human perceptions of fairness.

What it does

FairTik uses multi-modal AI and human-aligned reinforcement learning to ensure fair revenue sharing on TikTok. It processes videos and live streams, evaluates them with GPT-4o across multiple dimensions, and then optimizes revenue splits through fairness-aware alignment. The result: creators feel valued, and audiences can trust the process, and TikTok benefits from a more engaged, motivated community.

How we built it

Processed video content by sampling frames and generating transcripts to capture multi-modal signals.
Applied GPT-4o to score content quality across various dimensions.
Constructed a synthetic dataset of revenue-sharing proposals using the GPT-4 API.
Adopted Direct Preference Optimization (DPO) to align the model’s outputs with human fairness judgments.
Every revenue allocation includes a reasoned justification, so creators understand the rationale.
Leveraged Lynx for cross-platform interfaces, applied context engineering to improve AI responses, generated structured data from unstructured inputs, and used AI-powered code editor to automate development.

Challenges we ran into

No existing dataset for fairness alignment, so we had to design a data generation pipeline from scratch.
DPO is sensitive to data quality, making it critical to generate synthetic datasets that are both diverse and high-quality.
Rapidly learning and implementing the Lynx framework posed challenges, particularly due to its limited ecosystem, requiring us to solve visual design and cross-platform issues from first principles.

Accomplishments that we’re proud of

Designing a complete end-to-end pipeline from video evaluation to fairness-aware revenue sharing.
Successfully generating synthetic fairness-aligned data for training.
After DPO fine-tuning, the model can output reasonable reward distributions along with convincing justifications.
Demonstrated that combining AI evaluation with human-aligned reinforcement learning can produce fair, trustworthy outcomes.
Delivered an innovative UI that integrates chatbot and intelligent navigation, featuring modern gradients and seamless cross-device experiences via Lynx.

What we learned

How to apply DPO to a real-world fairness problem.
Practical experience in building synthetic datasets with large language models.
The importance of combining automated evaluation with human-aligned preferences to achieve fairness.
How natural language-driven interfaces (chatbot + intelligent navigation) can significantly improve user experience on complex data platforms.

What’s next for FairTik

Acquire real, high-quality datasets for post-training and evaluation.
Explore partnerships with platforms to integrate FairTik as a transparent reward engine.