Perception team with focus on model evaluation for onboard models and foundation models.

TikTok Search Video Retrieval and Ranking (leading video indexing)

• Introduced LLM to search indexing for for measuring video quality to improve content quality.

• Improved search indexing by introducing OCR and adding new prediction head to the model. Improved retrieval by coverage by 20% in experimentation.

• Redesigned offline indexing pipeline primarily by migrating from MapReduce to PySpark, reducing runtime of pipeline from 30 hours to 4 hours and reducing max cpu cores needed from 60k to 15k.

TA for ML courses under Dr. Ruslan Salakhutdinov, Dr. Eric Nyberg, and Dr. Matt Gormley.

• Worked in the Trust & Safety team.

• Improved the accuracy of pornographic audio detection model by 15%.

• Updated TikTok codebase to support self-suprivsed learning models and self-supervised model training (e.g. Data2vec).

• Develop initial multimodal model for fusion of video and audio backend models.

SWE internship focusing on computer vision algorithm development.