Erishen Sun

Open Source · 2026/08 — Present

• Original AI Agent toolchain: resolve-harness (Python / LangGraph agent skeleton), resolve-studio / resolve-tui (TypeScript / Rust runtimes), covering Agent Harness, Skills, MCP, Memory and Multi-Agent orchestration as a cross-language reusable Agent ecosystem.
• Original Fast-path plugin runtime: strips deterministic problems away from the LLM (built-in matchers with zero model calls on hit → codegen detectors validated by an AST-whitelist sandbox → well-behaved detectors promoted into source), significantly reducing model call costs.
• resolve-skills: a language-agnostic Agent Skills open standard (compatible with Claude Code / Codex) unifying the runtimes; distilled the PSE (Planner / Specialist / Evaluator) verification collaboration pattern.

Open Source · 2026/08 — Present

• Async RAG document pipeline in Rust / axum: document upload, chunking, Embedding, vector write and hybrid retrieval (BM25 + vector) in one service.
• Seven-layer architecture: from storage (SQLite exact dedup + perceptual hash) through retrieval, LLM re-ranking to generation, decoupled per layer; wiring local index and LLM re-ranking into an evolvable pipeline.
• Local-first and privacy-oriented: index independent of the file tree, incrementally re-scannable; classic algorithms + heuristics over heavy ML deps; paired architecture write-up open-sourced.

Open Source · 2026/08 — Present

• AI BI workbench: connect SQLite / CSV / PostgreSQL / MySQL, ask in natural language, LLM writes read-only SQL and renders ECharts dashboards.
• Deterministic Text2SQL pipeline (schema constraints + read-only execution + result validation) runs in parallel with a ReAct dual-path (tool calls + reflection/retry), balancing control and flexible exploration.
• Read-only and auditable by design: pre-execution permission / scope checks, errors located and surfaced; built for self-service BI; paired architecture write-up open-sourced.

Open Source · 2026/09 — Present

• Full-stack AI Agent development framework on Spring AI Alibaba (Alibaba Cloud Bailian / DashScope, direct access in China): Chat / RAG knowledge base / ReAct Agent / long-task interaction modes, Java 17 / Spring Boot 3.5 backend + Vue 3 frontend.
• Planner → Specialist → Evaluator three-role orchestration (parallel subtasks, failure retry, overall review); RAG knowledge base (upload → embedding → Rerank retrieval) generalized as a unified tool across modes.
• 8 local tools + MCP toolchain; Docker code sandbox isolating 8 languages (network isolation / read-only filesystem / resource limits); long tasks persisted in SQLite with resumable interruption; cross-session long-term memory (optional AES-256-GCM encryption); API keys never leave the backend, MCP output auto-redacted.

PayPal · 2026/01 — 2026/07

• Test environment engineering: built 20+ local test environments (16 from scratch) covering config, ordering, redirect, callback, order status, authorization, capture and refund; unified start / reset / test / verify flows via Docker Compose / Makefile / Playwright.
• AI verification process: landed the Planner / Specialist / Evaluator (PSE) three-role verification loop in the payment-migration project — Planner defines tasks and acceptance criteria, Specialist executes migrations, Evaluator verifies results against the criteria — converging AI-assisted development output into a reproducible, auditable engineering process.
• API evidence analysis: empirically validated 34 Orders V1/V2 fields, identifying 18 (~53%) documented-but-unenforced constraints, forming an API_ENFORCED / DOCS_ONLY taxonomy.
• Migration delivery: integrated and tested legacy NVP / SOAP payments across multilingual e-commerce environments, delivering verification reports combining API-behavior evidence, migration rules and Playwright E2E tests; fed defects and fix proposals back to the migration tooling team.
• Agent practice: built an AutoGen-based payment migration Agent (file I/O + PayPal API + Playwright E2E + FastAPI + SSE log stream + Web UI) to validate agent-driven payment migration and engineering verification flows.

PayPal · 2025/01 — 2025/12

• 0→1 architecture & delivery: led frontend and BFF-layer architecture and development of the CIP identity verification platform, owning the full delivery from design, development and testing through security audit to production release.
• Core business capability: implemented multi-vendor dual-channel liveness detection (Baidu / Tencent), supporting the Plan / Retrieve / Result core flow and traffic-ratio configuration.
• Security & compliance: built privacy policy management, XSS protection, SSO authentication and Redis session management around PIPL privacy compliance and enterprise security requirements.
• Observability: established Prometheus / Jaeger / Grafana monitoring and tracing, completing production delivery with the PayPal Pentest / SDER security audit.

Trip.com Group · 2020/01 — 2024/01

• Core content platform: owned React SSR architecture and iteration for the Blog / Destinations international content channels, delivering 14 pages and 100+ reusable components.
• Internationalization & SEO: supported Sitemap generation and localization for 10 languages x 185 regions, landed Schema.org structured data, and continuously improved international search indexing and traffic.
• Core business optimization: extended Next.js source with configurable resource-loading parameters, packaged as an internal npm module reused across projects; Blog H5 revenue ranked behind only flights and hotels.
• Performance engineering: owned Next.js upgrades, Webpack bundle splitting and SSR pipeline optimization; contributed to Core Web Vitals governance.
• Multi-platform delivery: built Trip App React Native pages and mobile video-playback compatibility for a consistent Web / App content experience.