AI Application Engineer · Shenzhen Linyuan Technology 2025.10 – 2025.12

Built a multimodal AI content generation system for cross-border e-commerce visual supply chains.

• Designed high-availability async architecture: FastAPI + Redis distributed task queue + WebSocket push + dead-letter retry — zero crashes under peak traffic
• Built LLM routing gateway with adapter pattern unifying Gemini, SD, and other heterogeneous models
• Constructed LLM-as-a-Judge evaluation benchmark with Few-shot + CoT prompt optimization — reduced manual intervention by 30%+

Twinbuddy — AI travel companion matching platform

Core moat: bilateral LangGraph avatar negotiation accumulates behavioral data (concession patterns, conflict styles) that competitors cannot replicate by swapping LLMs.

React + Vite FastAPI LangGraph PostgreSQL Redis Qdrant Claude API

Tencent × SZTU: AI Code Analysis & Multilingual Repair Benchmark 2026.03 – present

Built a high-difficulty code repair benchmark for internal AI Coding Agents.

• Fine-tuned Qwen3.5-9B with LoRA for domain-specific code analysis; built AST slicing + 3-way RRF fusion RAG pipeline
• Designed 4-dimensional system-level prompt optimization + adversarial dual-review workflow
• Cross-file code analysis F1: 0.28 → 0.61 (+120%)
• Designed C/E/Q 3-dimensional continuous scoring + regression penalty — surpassing SWE-bench binary scoring limits
• Built fail-to-pass test suite covering 5 languages × 3 defect dimensions from scratch

obsidian-shared-memory-bus — Cross-agent shared memory 2026.01 – present

Solves context isolation across AI coding tools (Claude Code, Codex, Cursor, OpenCode).

• L0–L4 five-layer memory architecture: Event buffer → Durable persistence
• BM25 + Dense hybrid retrieval + FNV-1a32 LSH offline vectorization — <50ms latency on 100k-token local libraries
• MMR (λ=0.7) diversity reranking + Circuit Breaker + Backpressure — stable at 50+ QPS

LangChain LangGraph Multi-Agent RAG LoRA MCP BM25 WebSocket Prompt Engineering