Erishen Sun
AI-native Engineer | Agent Engineering · AI Code Verification · React / Node.js Full-stack Delivery
erishen.cn+86 186****6513+86 186****6513WeChat: erishenerishen [at] qq [dot] comerishen [at] qq [dot] com
Summary
- Delivering production-grade applications and verification systems with React / Node.js + AI agents, on years of full-stack production experience — and building tools from scratch rather than following tutorials: a C11 single-binary runtime to an agent toolchain, each end-to-end by me.
- Shipped three series — an AI verification framework, a deterministic runtime, and a single-binary backend — and keep writing (90+ technical articles in Chinese and English), each backed by runnable code and a write-up.
- I validate design trade-offs with runnable code, automated tests and independent acceptance rather than concept demos — and I keep exploring the engineering loop that takes AI-generated code from execution, through verification, to delivery.
Skills Matrix
AI Engineering
EngineeringPSE three-role verification loop, Fast-path deterministic executionAgentsHarness, Workflow, Skills, MCP, Memory, Multi-AgentAppsRAG, AI-assisted developmentFrameworksAutoGen, LangGraph, CrewAI
Engineering
DeliveryDocker, CI/CD, Playwright, automated testingObservabilityPrometheus, GrafanaSecurityKYC / PIPL, privacy compliance, security verification
Frontend
FrameworksReact, Next.js (SSR / App Router)Cross-platformReact Native, Taro, Tauri 2, ElectronGraphicsThree.js / WebGL, real-time communication
Backend
LanguagesTypeScript, Node.js, Python / FastAPI, Java / SpringSystemsRust, Go, C / ShellStorageMySQL, Redis
Work Experience
PayPal
Full-Stack Engineer
Tech Stack: TypeScript, React, Next.js, Node.js, Express, Docker, Playwright, MCP, Claude Code Skills, Redis, gRPC, Jest, Vitest, Makefile, Python, FastAPI
- Full-stack ownership across three tracks at PayPal China: enterprise platform frontend and BFF architecture, payment-integration migration engineering, and AI-assisted development flows.
- Led the CIP identity verification platform frontend and BFF layer 0→1, with the project passing the relevant penetration testing (Pentest) and security-design review.
- Owned the COS customer operations platform frontend: a gRPC debug tool cut issue localization from 1-2 hours to ~10 minutes, KYC pre-validation cut invalid operations 40%, and 10+ production releases shipped.
- Payment upgrade (NVP / SOAP → REST): Skills-driven migration, independently verified across 20+ environments and 60+ E2E tests (implementation detail in the "Projects" payment-migration card).
- Built PayPal China official-site and Help Center pages: cut the per-page dev cycle from ~6 hours to ~1.5 hours (~4x); see the "Projects" official-site card for the AI-assisted dev method.
- Engineering delivery: Next.js 12→14 and Node.js 14→20 upgrades across 4 production projects, raising core-module unit-test coverage from 40% to 75%, with 30+ production releases and zero P0/P1.
Shanghai Quante Information Technology
Full-Stack Engineer
Tech Stack: React, Next.js, Three.js, Node.js, Socket.io, Azure AI
- Led frontend development of a Disney Resort AI digital-human customer-service system, validating AI + 3D interaction in a real deployment.
- Implemented web 3D digital-human rendering with lip / expression sync using react-three-fiber; integrated dual-channel Azure ASR / TTS voice with real-time interruption.
- Resolved iOS / Android mobile WebGL and audio compatibility issues, keeping the experience consistent across real devices.
Trip.com Group
Senior Frontend Engineer
Tech Stack: React, Next.js, Node.js, Express, SSR, GraphQL, Serverless, Redis, MySQL, React Native, Taro
- Owned the Trip.com international content platform; built the department’s first Node.js SSR + BFF content platform, distilling an internal npm package and an SSR asset-loading strategy.
- Delivered 14 core pages and 100+ reusable components; contributed to Core Web Vitals optimization.
- Delivered live-streaming H5, community features (RN / H5 / mini-program) and rich-text editing across the full content pipeline; led a 2-4 person frontend team.
Early Career (2005-2017)
Software Engineer → Technical Manager → Frontend Development Manager (Tech Lead)
Tech Stack: Java, Linux/C, Hybrid, XMPP, React Native, Jenkins
- Started in Java enterprise backend (3G NMS / CRM), through Linux / C router firmware, hybrid mobile apps (XMPP IM), to React Native cross-platform apps.
- Progressed from Software Engineer to Technical Manager to Frontend Development Manager (Tech Lead), leading 3-7 person teams; collaborated across China-US time zones (Actiontec).
Projects
Selected Work Projects
PayPal — AI-Driven Payment Integration Migration Engineering
PayPal · 2026/01 — 2026/07
- Merchant-facing payment integration migration (NVP / SOAP → REST): unified "start - reset - test - verify" flow on Docker Compose / Makefile / Playwright, covering checkout, redirect, authorization and refund.
- Designed a three-layer image cache (deps → build → app) that solved arm64 / amd64 native-extension incompatibility in one place and supports offline cold start.
- Developed payment-migration Skills: detect stack and legacy API, route to per-stack expert Skill, rewire the legacy entry to a team-built Order V2 one-shot module copied into the client env.
- Traced repeated divergent architectures from one Skill back to a design gap (weak instruction constraints); improved execution consistency through explicit constraints, structured output and independent acceptance.
- Personal verification tool (not a company initiative): practiced and distilled the PSE (Planner–Specialist–Evaluator) method during payment-migration verification — separating task planning, AI execution and independent acceptance into a reproducible, auditable verification loop; paired with an AutoGen migration agent (Playwright + FastAPI/SSE + Web UI) that end-to-ends "pick env → trigger migration → live results".
PayPal — China Official Site & Help Center
PayPal · 2025/03 — 2025/12
- AI Figma-to-Code: Figma MCP extracts design specs without manual measurement, Serena MCP does semantic search and symbol-level edits, Playwright MCP screenshots verify design fidelity; 20+ docs captured.
- Built China official-site landing pages and Help Center pages, shipped to real users: a CBT Carousel Card (Swiper + dynamic offset math) and a Layered Card (brand restyle), responsive across breakpoints.
- Merchant Stories cross-border brand pages and overseas-market landing pages with multi-language switching; Help Center search hardening, mobile adaptation and Livechat toolbar integration.
Quante — Disney Resort AI Digital-Human Concierge
Shanghai Quante Information Technology · 2024/01 — 2024/06
- Digital-human interaction: dual-channel Azure ASR / TTS and FunASR over Socket.io, linking speech and interruption to the avatar’s speech / motion / expression sync, with four-layer interruption design.
- Lip-sync and avatar rendering: evaluated BlendShape approaches (Azure standard / NVIDIA Audio2Face / OVR LipSync) on accuracy vs latency, mapping phonemes to 3D facial animation; integrated howler.js.
- Built the knowledge-base admin console (Next.js + Express + MySQL): knowledge-base CRUD, Q&A maintenance and semantic-matching configuration, based on the open-source nsgm-cli scaffold.
Open-Source AI Engineering
Resolve — AI Agent Toolchain & Fast-path Runtime
Open Source · 2026/08 — Present
- An agent toolchain covering Harness, Skills, MCP, Memory and multi-agent workflows.
- Fast-path is the core design: deterministic tasks cost zero model calls; on a miss the LLM emits a detector that is AST-whitelist-sandboxed, verified and persisted, so later requests of its kind run directly.
- PSE (Planner–Specialist–Evaluator) three-role verification: generation and acceptance are split, and the evaluator holds no context from the generator, which targets self-checking (sycophantic) hallucination — making output trustworthiness a property you can verify, not vibes.
Lume — C11 Single-Binary Backend + Agent Scripting Language
Lume GitHub (host · agent-httpd full)Lume language core (3 backends / AI self-verify)Live · lume-crm (< 20 MB deployment proof)
Open Source · 2026/10 — Present
- Core implementation: C11 single-binary backend (~1.7 MB) serving HTTP / JSON / SSE / LLM tools / SQLite; < 20 MB total image, edge-deployable.
- Verification: in progress — the same .lume program runs identically across interpreter / LLVM-IR / libLLVM backends, with fuzzing exploring repeatable, verifiable code execution.
- Goal: let AI-generated code run, replay and verify inside a constrained runtime — an acceptance base for AI code.
Education
Nanjing University of Information Science and Technology
Bachelor's Degree · Information and Computing Science
2001/09 — 2005/06