Worked on data and training optimization for Seed-Omni, spanning continued pre-training and mid-training across speech, text, and vision. Built a three-stage data-governance pipeline combining rule-based filtering, LLM-assisted review, and modality validation, together with sample- and token-level diagnostics for tracing loss spikes, data conflicts, and capability regressions. Developed reusable production and synthesis pipelines for dialect speech, VS2T, audio captioning and QA, SpeechQA, and long-form interleaved data, improving key workflows by up to 50×. Took end-to-end ownership of a 12B/128K, 463B-token mid-training run—from data recipe and training configuration through monitoring, checkpoint selection, evaluation, and regression attribution—and distilled the findings into reusable team practices.

Led EcoMLLM, a unified multimodal-LLM framework for large-scale short-video content understanding that replaced a stack of task-specific models. Designed a two-stage cascade—a multi-view VLM producing structured descriptions plus a compact LLM for fast downstream inference—and built Juggle Pipeline, a Ray + Merlin data system with 30+ operators for constructing large-scale instruction data. Combined training-free token pruning, FP8 quantization, and visual-token compression to increase QPS from 5.65 to 17.1 (>3×) without measurable downstream degradation. The system achieved 59.58% video-VQA accuracy and around 90% precision and recall on core tasks; RL (GRPO) with CoT cold-start and composite rewards improved US recall by 35% and instruction following to 100%.

Rank 1 / 191 · GPA 4.13 / 4.50 · CET-4 632, CET-6 551