Agent / AI Engineering · CS (创新人才实验班) @ Guangzhou College of Commerce · 2023–2027
Building AI-native full-stack systems — from streaming runtimes & Agent event channels to security-hardened backend integrations.
- Champion, Security Track — Anker Innovations 1st Hackathon Challenge (安克创新首届黑客松挑战赛 · 安防赛道冠军)
2026 - Champion, AI³ Growth Journey Season 8
2026 - Champion, ConsensusHK
2026 - National Top 40, 2026 AI Pioneer Future Talent Competition (2026 AI 先锋未来人才大赛)
2026 - National First Prize, National College Students' "Innovation, Creativity & Entrepreneurship" E-Commerce Challenge (全国大学生电子商务「创新、创意及创业」挑战赛)
2026 - Provincial First Prize, Guangdong "Maker Cube Robot" Competition (广东省创客魔方机器人比赛 · 高中组)
高中
AI Full-Stack Intern · Baidu (China) Co., Ltd. Guangzhou Branch 2026.07 – Present
GenFlow (pnpm monorepo multi-device AI framework) runtime + AICRM customer-data pipeline.
GenFlow — 多端 AI 应用框架(AI 对话 / PPT 生成 · 一次编写、四端交付)
- Unified streaming runtime: business calls
useSseStream(runtime).start({prompt}); the platform package injects SSE (PCEventSource/ NA JSBridge duplex / WAP weak-network fallback) and normalizes raw frames before handing off to message components - Agent-oriented event channel:
runtime.eventBusdecouples SSE producers from message/state/debug consumers; plugins register viaruntime.use(plugin)on the Chat/PPT lifecycle - Cross-platform isolation: shared logic sinks to the common layer with per-platform providers absorbing device diffs; dual vue2/vue3 entries switch by preset; playground mocks the AI stream for 4-device side-by-side debugging
AICRM — 企微 × One CRM 客户主数据打通 · 加好友事件驱动链路
销服 BI → dbt 数仓 → FastAPI(/v1 / /skill / /internal)→ Web + AI 技能; owned the customer master-data & event-integration backend- Deterministic matching (phone → company full/short name) under a "never mis-bind" policy; live
getWeChatRelre-check before write (cache never used as write source); 3/hour + 30-min cooldown - Idempotent add-friend events (202 async) with
FOR UPDATE SKIP LOCKED+ 120s-lease workers; snapshot table + scheduled refresh cutGET /customers32s → 0.1–0.3s
AI/LLM Engineering Intern · Guangzhou Xin'an Data Co., Ltd. 2026.04 – 2026.07
Built Multi-Agent document review pipelines for enterprise brand & KPI compliance systems.
- Preprocessed stamped scans (OpenCV de-stamp/de-watermark + CLAHE) before
PaddleOCR; fused title (0.6) + full-text (0.4) vectors; hierarchical Markdown chunking + Hybrid Search + Reranker onMilvus— QA accuracy 60% → 88% on a 1,000-metric eval set - Built the "review → suggestion → revision" loop: Dify multi-node workflow, 7 intent-routed parallel retrieval, 5-element structured verdicts; similarity-aided report ingestion to avoid duplicate archives
- Provider-agnostic async LLM connection pool (rate-limit, exponential backoff, retry, streaming); multi-arch ARM64 images via
docker buildxacross 13 services withrequest_id + task_idtracing
Founder / Full-Stack Developer · Guangzhou Kangxiaomiao Smart Technology Co., Ltd. (BoBaCarry) 2025.09 – 2026.04
Built a shared smart-follow-cart platform end-to-end, from WeChat Mini Program to production deployment.
Node.js + Expressbackend withMySQL + Redis(<50ms response latency);Vue 3 + Element Plusops console; WeChat Mini Program (device binding / live status / history / login)- Defined MQTT device protocol + built a device simulator before hardware arrived; fixed a reconnect-without-resubscribe bug; virtual queue + structured concurrency for batch dispatch
- Intent-routed customer-service loop (query / urge / repair) with tool calls; token-cost attribution via
LangSmithtrace normalization; Nginx + HTTPS + PM2 + COS with circuit-breaker & idempotency
Contributing across the AI/agent ecosystem — from MCP servers and the Vercel AI SDK's frontend components to Python event-bus internals and cross-platform CI correctness.
| Repo | Type | Status |
|---|---|---|
| line/line-bot-mcp-server #590 | Feat: add get_group_summary MCP tool |
✅ Merged |
| NandhaKishorM/laya #216 | Fix: cross-platform encoding/CRLF bug in CI benchmark audit | ✅ Merged |
| browser-use/bubus #35 | Feat: EventBus.remove() to unsubscribe handlers |
In review |
| browser-use/jev-ultrafast #188 | Fix: browser-agent process leak after recording | In review |
| HKUDS/nanobot #5729 | Fix: deterministic hash for WeCom/WeChat media fallback names | In review |
| vercel/ai-elements #476 | Fix: overflow-hidden → overflow-x-clip (sticky positioning) |
In review |
| vercel/ai-elements #477 | Fix: decouple ImageProps from deprecated SDK type |
In review |
| vercel/ai-elements #478 | Fix: Enter-key guard bypass during streaming | In review |
| gokite-ai/kite-x402-services #2 | Feat: add Frankfurter FX x402 service | In review |
| vercel/ai #10082 | Discussion: proposed token-aware batching for embedMany |
Closed |
ParTrace Vision 驿站追迹 — AI event-tracing agent for parcel stations 2026.10
🏆 Champion, Security Track — Anker Innovations 1st Hackathon. Android (Kotlin + Compose) + FastAPI + Agent package + Vision worker, PostgreSQL + pgvector, deployed on AWS.
- Anti-hallucination delivery: agent replies ship with deterministic cards assembled only from verified tool payloads; SSE run-event stream with snapshot-first +
Last-Event-IDresume - Passwordless Aurora IAM auth: stdlib SigV4 presigning byte-identical to
aws rds generate-db-auth-token, IMDSv2 role credentials, zero new dependencies - Spec-driven development with OpenSpec: 11 tables / 27 APIs aligned across modules, SQL, and OpenAPI, guarded by a docs-consistency CI
Python FastAPI PostgreSQL pgvector AWS Kotlin
LingMate 聆伴 — AI voice life assistant for the elderly (Age Well Hackathon) 2026.08
LangGraphStateGraph (understand → execute → respond) routing to 5 single-purpose sub-agents with slot filling and natural-language time parsing- Three-layer anti-fraud detection: weighted rule engine + RAG similar-case retrieval + LLM review; rules keep working when the model is down (34/34 unit tests)
- Voice pipeline: Vosk offline ASR + Tencent Cloud ASR/TTS with graceful fallbacks; OpenAI-compatible LLM layer over Hunyuan / DeepSeek
Python LangGraph FastAPI React 19
HUM / Voices 人声余温 — Emotional voice-persona platform 2026.01 – 2026.02
🏆 Winner, AI³ Growth Journey Season 8. Each Voice = timbre (StepFun API cloning) + personality + emotional-boundary strategy.
- SSE-streamed TTS with React state + Audio Context double-buffering — <300ms time-to-first-audio
- Redis semantic/audio caching cut repeated-dialogue TTS cost by 45%
- On-chain persona provenance on
0G(AI + Data native chain)
TypeScript React Node.js PostgreSQL Redis Docker StepFun API
EdTech Platform — Adaptive learning path generation 2025.11 – 2026.01
- Extended
BKTcognitive diagnosis model with Redis-backed real-time mastery-probability updates - Few-shot-injected BKT mastery state into
Gemini APIfor difficulty-calibrated question generation
Spring Boot Gemini API React RabbitMQ MyBatis Plus
indicator_detector — KPI quality detection & recommendation 2026.04 – 2026.05
- Hybrid scalar-filter + vector-similarity retrieval on
Milvus Litefor multi-tenant Top-K KPI recommendation - ReAct-style prompt design (Few-shot + CoT) for formula-dependency parsing with self-correcting JSON Schema output
Python Milvus Lite Transformers PyTorch
Multi-Agent Orchestration (Dify / LangGraph) MCP RAG (Hybrid Search + Rerank) Prompt Engineering (CoT/Few-shot) LLM Connection Pooling JWT/RBAC SSO / Identity Gateway
154+ contributions this year · 23 public repos · Pull Shark 🦈

