Dec 2025 – Mar 2026 · Beijing

ByteDance · Flow

Model Data Platform Product Manager

Supported expert annotation for Seed Code post-training across the full data pipeline: trajectory generation, interactive correction, quality review, and feedback. I contributed to the interactive annotation platform so experts could correct model reasoning, tool calls, and execution results step by step. I also kept the workflow aligned with emerging Agent behavior and used AI-assisted checks to improve both annotation efficiency and data quality. The core idea was to capture expert judgment efficiently and reproducibly as training data that genuinely improves model capability.

40%Issues resolved during annotation by AI QAInteractive Agent trajectory annotation platformAligned with new MCP / Plan capabilities

My role

Contributed to the interactive trajectory annotation platform, translated algorithm-team toolchain requirements into product capabilities, and kept annotation queues and QA rules aligned with evolving Agent behavior.

Interactive annotation design

  • Helped design an interactive trajectory annotation platform around three core elements: Agent reasoning, tool calls, and execution results, with step-by-step expert correction.
  • Supported inserting or editing intermediate steps without regenerating every later step, reducing annotation cost.

Agent capability alignment

  • Worked with algorithm teams to support emerging patterns including multiple tool calls in one step, internal subagent trajectory annotation, and Plan.
  • Updated annotation guidelines and QA rules in parallel so training data stayed aligned with the latest Agent behavior.

AI-assisted quality assurance

  • Evaluated machine-QA capabilities and, given their high-precision, low-recall profile, designed a two-path approach: pre-checks during annotation plus AI assistance during quality review.
  • Moved issue detection earlier in the workflow; around 40% of issues were resolved during annotation, reducing rework and handling time.

What I learned

Having the model generate first and experts correct its trajectory—rather than feeding it only ideal answers—covers the states the model actually reaches. A data platform makes that expert judgment stable, repeatable, and reusable.