Beijing, China · On-site

Lead evaluation-driven Code Agent post-training for WorkBuddy, connecting capability diagnosis, targeted training data, and reinforcement learning. Built separate functional and aesthetic evaluation, difficulty-based data selection, and quality control for outputs and teacher trajectories; integrated SGLang with browser sandboxes and developed probability-aware GRPO clipping for low-probability tool-call tokens. Improved expert agreement to 90.7% and reduced training time by approximately 74% in the reported comparison. Browser-based verification also accelerated reward computation, and the evaluation pipeline now supports model iteration and training decisions, connecting real user needs, capability diagnosis, and model training to enable continuous evolution through an iterative feedback loop.

Shenzhen, China · On-site

Built a one-million-pair multilingual translation corpus for QQ International spanning 13 Eurasian languages, covering data filtering and cleaning, low-resource language-pair distillation, NER entity repair, 4B-class base-model evaluation, and SFT. The resulting model improved in-domain performance by approximately 10 percentage points while matching a substantially larger MoE translation baseline on FLORES-200 and WMT25.