Experience

Timeline of roles — enterprise AI, agent systems and forward deployment.

Agent Developer

Current · Beijing

Kuaishou · Beijing, China

Aug 2026 – Present

Responsible for the evaluation system, Figma structured editing, and MCP integration of M0 Design (PRD → Figma → Code) Agent at Main Site Technology (Operations Tech).

Led continuous evaluation platform for PRD → Figma Agent, constructing a benchmark of 84 tasks / 160 pages / 979 atomic criteria; combined Figma AST parsing, screenshots, Rule-based Graders, and LLM-as-a-Judge, achieving 95%+ core acceptance coverage, dropping single-case eval time from 20-30m to 3-5m, and cutting high-priority bad cases from 18% to 6%.
Co-designed declarative editing protocol serializing Figma SceneGraph into JSX structures with Reconcile diff writeback; achieved 95%+ complex design edit success rate on 143+ test cases, reducing fine-grained tool calls by 60%+.
Packaged generation capabilities into MCP for internal agents (Agent → MCP → Figma Runtime / Host); governed timeouts and runtime crashes, achieving 95%+ E2E success rate and halving concurrency exception rates.

View 3 contributions

Led continuous evaluation platform for PRD → Figma Agent, constructing a benchmark of 84 tasks / 160 pages / 979 atomic criteria; combined Figma AST parsing, screenshots, Rule-based Graders, and LLM-as-a-Judge, achieving 95%+ core acceptance coverage, dropping single-case eval time from 20-30m to 3-5m, and cutting high-priority bad cases from 18% to 6%.
Co-designed declarative editing protocol serializing Figma SceneGraph into JSX structures with Reconcile diff writeback; achieved 95%+ complex design edit success rate on 143+ test cases, reducing fine-grained tool calls by 60%+.
Packaged generation capabilities into MCP for internal agents (Agent → MCP → Figma Runtime / Host); governed timeouts and runtime crashes, achieving 95%+ E2E success rate and halving concurrency exception rates.

View Kuaishou work details & methodology

Agent SystemsFigma AST / JSXEvaluation BenchmarkMCPPythonTypeScript

AI Application Engineer

Bananain (三立人科技) · Shenzhen, China

Apr – Aug 2026

Led internal AIGC visual production and enterprise Agent platform engineering and governance in the digital R&D group.

Led a product image generation platform and PS AI plugin, designing LookSpec, DAG orchestration, BBox local editing and quality retries. ROI projections targeted 10→30 images per person per day and ¥89.8→¥43.9 per image (−51%); actual production outcomes remain unverified.
Built an AI Gateway and MCP capability layer for model, data and business API access, with identity and permission scope. Advanced data integration and training preparation for the AI Planning Agent; subsequent adoption remains unverified.

View 2 contributions

Led a product image generation platform and PS AI plugin, designing LookSpec, DAG orchestration, BBox local editing and quality retries. ROI projections targeted 10→30 images per person per day and ¥89.8→¥43.9 per image (−51%); actual production outcomes remain unverified.
Built an AI Gateway and MCP capability layer for model, data and business API access, with identity and permission scope. Advanced data integration and training preparation for the AI Planning Agent; subsequent adoption remains unverified.

FastAPIReactAdobe UXPMCPAgent PlatformRAG

Agent Evals & Code Intelligence Lead

Collab · Remote

Tencent × SZTU · Remote (University Collaboration)

Mar – Sep 2026

Tencent mini project · RepoMind Code Intelligence & Agent Quality Evaluation: Multilingual code-intelligence and quality evaluation system for coding agents.

Built replayable Observation Kernel across 496 real agent traces; applied Replay, Grounding Audit, and Guardrails; staged evaluation across Evidence Retrieval / Reasoning / Patch Validation, improving cross-file F1 from 0.28 to 0.61.
Built Tree-sitter multilingual code graph and incremental indexing with snapshots, immutable generations, and Freshness Gate; reduced single-file update latency from 10.78s to 0.12s (~90×) across 310 files / 3,855 symbols.

View 2 contributions

Built replayable Observation Kernel across 496 real agent traces; applied Replay, Grounding Audit, and Guardrails; staged evaluation across Evidence Retrieval / Reasoning / Patch Validation, improving cross-file F1 from 0.28 to 0.61.
Built Tree-sitter multilingual code graph and incremental indexing with snapshots, immutable generations, and Freshness Gate; reduced single-file update latency from 10.78s to 0.12s (~90×) across 310 files / 3,855 symbols.

MCPCode GraphTree-sitterASTTrace & ReplayEval

AI Application Engineer

深圳市麟远科技 · Shenzhen, China

Oct 2025 – Feb 2026

Built an e-commerce visual generation pipeline with FastAPI + Redis + WebSocket, unifying multi-model routing.

Competitor discovery → VLM visual-style decomposition → prompt/style library.
FastAPI + Redis + WebSocket async pipeline; LLM-as-a-Judge evaluation.
Cut API cost ~20% and manual review/rework ~30%.

View 3 contributions

Competitor discovery → VLM visual-style decomposition → prompt/style library.
FastAPI + Redis + WebSocket async pipeline; LLM-as-a-Judge evaluation.
Cut API cost ~20% and manual review/rework ~30%.

FastAPIRedisWebSocketVLMMulti-Model RoutingLLM-as-a-Judge

Team Lead · Product & Engineering

TwinBuddy (拼途旅行) · Shenzhen, China

Jan – Jun 2026

Travel-companion matching through digital personas and dual-agent negotiation. Track 3 Second Prize at the Qingbei-Harbin station of the 2026 Douyin AI Changemakers Hackathon League; competition completed.

MING persona modeling and LangGraph bilateral negotiation.
FastAPI + React / TypeScript prototype with explainable matching results.
Led the team and the main product / engineering work; development has ended, and the project is retained as an engineering case study.

View 3 contributions

MING persona modeling and LangGraph bilateral negotiation.
FastAPI + React / TypeScript prototype with explainable matching results.
Led the team and the main product / engineering work; development has ended, and the project is retained as an engineering case study.

LangGraphMulti-AgentPersonaFastAPIReactProduct