Experience

Tencent FiT

Post-Training Algorithm Engineer
Shenzhen, China
2026 - Present

Algorithm Engineer Intern
July 2025 - 2026

Main model contributor to NeoData, which turns user queries into retrieval over large-scale licensed financial data through financial entity extraction and MCP tool calls.
Serves tens of millions of daily active users; the sole financial data provider for Yuanbao, one of the main financial data interfaces for WorkBuddy, and a data backbone for many other financial businesses.
Iterate on multiple small models along the pipeline, covering financial entity retrieval and ranking, and financial tool selection and calling (tool choice and argument filling).

One of the main contributors. In the second half of 2025, before mature agent training frameworks were available, extended verl to support multi-turn rollouts with tool calls and trained the agent end to end.
Built the agent around a Qwen3-30B-A3B policy and 79 MCP tools, covering entity resolution, multi-intent function calling, dependency-aware planning, and recovery from empty results.
Built three generations of training data: filtered 700K production queries into 1K multi-hop seeds, explored DAG-first generation, and converged on template-guided synthesis grounded in real user phrasing.
Trained with multi-turn SFT, rejection fine-tuning, and process-aware RL using answer-, DAG-, trajectory-, turn-, and parameter-level feedback.
Designed asynchronous query-aware compression for 128K rollouts, keeping the newest tool response intact while compressing prior observations off the rollout critical path.
Reached parity with DeepSeek V3.2 on internal benchmarks and surpassed it on FinSearchComp under the same business tool setting.

Algorithm Engineer Intern
Nanjing, China
December 2024 - May 2025

Worked on the team’s DeepSeek-R1 reproduction. Long RL runs suffered training collapse: reward dropped sharply mid-training and the model began producing incoherent output.
Introduced DAPO into the team’s internal RL training framework to address the collapse.
Added a dynamic sampler that prioritizes prompts with high reward variance, so each batch carries more learning signal; this extended the number of stable training steps by about one third.