CV

Staff Applied Scientist, Business Integrity

TikTok Pte. Ltd.

I lead a team of 7 full-time scientists and engineers building the AI foundation for TikTok Business Integrity, with a focus on large-scale foundation model data, moderation intelligence, and agentic system evaluation.

• Agentic ads moderation system. Built an agentic content moderation system from scratch, end to end: problem definition, data curation, agentic SFT, agentic RL and serving. Adaptive feature selection through tool calls improves recognition performance and adapts quickly to policy changes, at a throughput of more than 2 QPS per NVIDIA H20.
• Large-scale synthetic data for foundation models. Built a fine-grained integrity knowledge graph of more than 300 entities and an online multimodal synthetic-data pipeline that grounds real-world videos to granular risk concepts and generates aligned QA, captions, spatial grounding and agentic moderation tasks. The pipeline generates more than 2B tokens a day and continuously improves our in-house foundation model.
• Agentic system evaluation. Built large-scale evaluation infrastructure and benchmark datasets for internal risk-operation agents, with automatic multi-round conversations against a user simulator, sandboxed trajectory rollouts with more than 20 internal skills, and evaluation of long-horizon tasks and memory.

Staff Research Engineer

Gaussian Robotics

Built the technical foundation from scratch for machine learning model training, evaluation and deployment on low-compute robotics systems.

• Large-scale multi-task learning. Ran a dozen deep-learning recognition tasks on 2-TOPS edge devices and set up MLOps for fast model iteration.
• Model development toolkit. Infrastructure for self-supervised pretraining, supervised training, INT8 quantization (PTQ and QAT) and runtime conversion.
• End-to-end testing system. A hardware-in-the-loop (HIL) testing framework with a web interface for pre-release checks.
• Foundation model research. A semi-automatic annotation system based on the Segment Anything Model (SAM), with fine-tuning techniques for in-house data.

Senior Research Scientist, AI Lab

TikTok Pte. Ltd.

Research on novel view synthesis and large-scale visual search.

• Novel view synthesis. Co-developed MINE, dense 3D reconstruction from a single image with NeRF and multiplane images, published at ICCV 2021.
• Visual search for TikTok e-commerce. Built the visual search embedding model for large-scale product retrieval from TikTok videos and co-designed the search pipeline and re-ranking.

Senior Research Engineer

ViSenze Pte. Ltd.

Large-scale visual search for e-commerce.

• ICCV 2019 DeepFashion2 clothes retrieval challenge. Built the retrieval model and re-ranking pipeline that won 1st place.
• Image retrieval. Trained most of the company’s online retrieval models and built the team’s training codebase; moved all online models to efficient architectures (MobileNet, ShuffleNet) with no accuracy drop; developed multi-task embedding training for category relevance.
• Large-scale distributed training. Combined data and model parallelism to train arc-softmax embeddings over millions of classes.

Lead Research Engineer

Tuputech

Led a team of 5 research engineers on large-scale pedestrian recognition (Re-ID) and tracking.

• Person re-identification and tracking. Built a system for pedestrian recognition and tracking across multiple cameras, with a data-driven Re-ID development workflow.
• Image classification. Built and maintained a TensorFlow training platform that let solution engineers train and evaluate classifiers in a few clicks.

B.Sc. (Honours)

University of Glasgow

Computing Science

Pistis Technical Report

arXiv:2609.28554

TextSculptor: Training and Benchmarking Scene Text Editing

arXiv:2605.21090

A Deep-Learning-based System for Indoor Active Cleaning

IROS 2022

MINE: Towards Continuous Depth MPI with NeRF for Novel View Synthesis

ICCV 2021

1st place, ICCV 2019 DeepFashion2 Clothes Retrieval Challenge

ICCV 2019 DeepFashion2 Workshop