PAN SHIYU

1. Trust & Safety Process Architecture (Video & AIGC)

• End-to-End Moderation Pipeline: Architected and progressively refined moderation pipelines covering detection, risk tiering, enforcement, report, and appeal, built on unified infrastructure governing both safety moderation and labelling workflows across 8 markets (ID/VN/TH/PH/MY/BR/SG/TW).
• Policy Strategy & Governance: Formulated and standardised policies for content safety and labelling. Acted as the primary liaison to align cross-functional stakeholders (Business, Local, and Operations). Proactively adjusted and iterated policies in response to evolving business needs, emerging content trends, and regulatory shifts to ensure long-term platform health.
• AIGC Content Safety: Identified and catalogued potential risk categories inherent to AIGC content (e.g., religiously offensive content, IP infringement); designed and deployed ‘Moderation-before-Delivery’ pipelines ensuring all AI-generated images and videos are intercepted and reviewed before reaching users, safeguarding platform reputation from unvetted AIGC risk.
• Copyright Compliance & Audio Safety (Key Project): Owned the end-to-end development of an automated audio copyright detection system to mitigate platform legal risks and potential financial liabilities. Designed and implemented a closed-loop enforcement workflow, enabling automated actions such as video muting or content restriction upon detection of copyright violations, ensuring platform-wide compliance with intellectual property regulations.

2. Models Construction of Content Safety & Content

• Full Lifecycle Ownership: Governed conception, iteration, evaluation, and deployment for 10+ safety and content understanding models. Partnered cross-functionally with algorithm and engineering teams to translate policy requirements into model specifications and drive iterative performance improvement.
• LLM & Agent-Powered Moderation: Piloted and deployed LLM classification to resolve hard cases beyond traditional MOE model capability, improving accuracy by 20% over baseline. Designed an agent-based workflow to automatically collect, organise, and summarise moderation cases, generating structured training data to support continuous model iteration.
• Continuous Evaluation Loop: Established model-evaluation queues and machine-result QA process, instituting a continuous loop of deployment, evaluation, and recalibration.
• Model-Driven Automation Reducing Moderation Manpower: Elevated fully automation (model auto-labelling without human moderation) from full manual review to 90%+ automation rate across 8 markets, substantially reducing manual moderation manpower while ensuring safety risk was never compromised for efficiency gains.

3. Moderation & Operations Platform Development

• Moderation Platform and Iteration: Built and iterated the core moderation platform to improve reviewer efficiency and decision quality, including designing queues and interfaces based on the requirements of different tasks.
• Operation Platform Iteration: Built and iterated internal operations tools to improving manual operational efficiency across markets, including multilingual policy configuration, moderation results revision, quality arbitration, etc.
• Cross-Functional Delivery: Partnered closely with local teams, operations, business/marketing across 8 markets to gather requirements, align on priorities, and ship platform capabilities at scale. Used data-driven analysis of reviewer performance and operational pain points to inform roadmap prioritisation.

Responsible for industry research on single-disease quality control and reporting system.

Contribute to the sanction module construction of the TenPay anti-money laundering system.

Owned the construction and iteration of the PaddleHelix biocomputing platform.