AI engineering / platform engineering / senior full-stack engineer focused on landing AI in real engineering workflows.

Previously built customer-facing FinTech systems at PayPal with React, Next.js, and Node.js. Recent focus areas include AI-assisted engineering workflows, system security hardening, and large-scale framework upgrades.

I care less about one-off AI demos and more about systems that are runnable, reproducible, testable, and explainable.

🔍 Open to opportunities

• Landing AI in real software systems (AI-assisted engineering workflows)
• Agent workflow design: planning / execution / independent evaluation / evidence-driven iteration
• Quality & security engineering: reproducible workflows, deterministic checks, and security hardening
• Developer tooling: CLI workflows, prompt / memory systems, and MCP-based code context
• Full-stack systems with React, Next.js, Node.js, TypeScript, Python, FastAPI, Redis, and APIs

📍 Shanghai, China

🌐 Blog: https://erishen.cn

A single-binary C11 backend framework: business logic lives in .lume scripts, and one ~1.7 MB executable serves static sites, JSON APIs, SSE chat, LLM agent tools and SSR pages with native SQLite — no Node runtime, no nginx, no separate React backend. HTTP / chat / MCP come from the embedded agent-httpd submodule: a from-scratch HTTP/1.1 server (prefork event loop, CGI + FastCGI, built-in React SSR) whose agent layer is C as well — tool calls, skills, memory, ReAct and PSE loops — and which today drives real MCP servers, including a 170-tool reverse-engineering stack backed by a source-built Ghidra decompiler and an LSP code-context server. Exercised in production shape by lume-crm (a complete CRM: one binary + one .lume script), plus lume-lan, lume-nsgm and lume-talentlens.

The host-independent language core — lexer, parser, type checker, tree-walking interpreter — plus two native backends (an LLVM IR text emitter and a libLLVM C API emitter): the same source type-checks and runs identically on all three, guarded by cross-backend parity tests. The parser was hardened through ten instrumented libFuzzer campaigns that surfaced eight memory / stack bugs; each one is pinned by a regression fixture and patrolled in CI by a fuzz leg (deterministic seed replay + timed random phase + hang threshold), with ASan/LSan legs catching leaks that a macOS host cannot see at all.

A full AI Agent runtime and engineering workbench on the Cordis DI container — LLM backend / tools / Agent loop / approvals / skills / frontend are all plugins; switching models or adding tools is config-only. Built-in PSE (Planner/Specialist/Evaluator) three-role verification loop and harness orchestration, integrating Makefile scripts, MCP tools and OS-level sandbox (macOS Seatbelt / Linux bwrap). Part of the resolve-* toolchain: resolve-harness (Python/LangGraph skeleton with Fast-path Plugin Runtime — zero model calls for deterministic problems), resolve-tui (Rust CLI/TUI coding Agent), resolve-skills (Agent Skills open standard, consumable by Claude Code / Codex).

An async RAG document pipeline in Rust / axum: upload, chunking, embedding, vector write and hybrid retrieval (BM25 + vector) in one service. Metadata lives in SQLite (tasks / document metadata / events, WAL mode with optimistic concurrency), while the vector + BM25 backend is switchable — in-memory, local LanceDB, or PostgreSQL/pgvector with the two BM25 implementations kept schema-aligned. The design is written up as seven explicit decisions (WAL persistence, version-number concurrency, FNV-1a embedding hash, RRF fusion over weighted scoring, shared storage backend, LLM rerank with automatic fallback, resource-oriented Axum routing). Local-first and privacy-oriented, built on classic algorithms + heuristics rather than heavy ML dependencies.

An AI BI workbench: connect SQLite / CSV / PostgreSQL / MySQL, ask in natural language, LLM writes read-only SQL and renders ECharts dashboards. A 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 exploration; pre-execution permission / scope checks for self-service BI.

A debuggable C++ Electron digital human workbench — porting the firefly 3D virtual human (React + three.js / R3F) to Electron, with the main process loading C++ native modules via node-addon-api: audio energy / VAD / microphone capture / ASR / TTS wrappers. Integrates Whisper.cpp local offline speech recognition, Piper local speech synthesis, RNNoise noise suppression, and music playback; supports setting breakpoints and debugging C++ directly in VSCode (Node debug main process + lldb attach).

I write about AI engineering, agent frameworks, and full-stack practice on my blog: erishen.cn.

• Evolution of a Distributed Order System: How Spring Boot + React Survived High Concurrency from a Monolith — The step-by-step path from a monolithic order system to a distributed one.
• Text2SQL Deterministic Pipeline + ReAct Dual-Path: Architecture Breakdown of datapulse AI BI Workbench — Balancing controllability and exploration in an AI BI studio.
• An Agent Runtime Built on Spring AI Alibaba: Five Real Engineering Problems and Trade-offs — Chat / RAG / ReAct / long-running tasks, and what each decision cost.
• From SQLite to LLM Reranking: Seven Design Decisions for a Rust RAG Service — The seven decisions behind the async RAG document pipeline.
• cicdkit Engineering Practice: A Local CI/CD Single-Binary Platform Built with Pure Go Standard Library — A local CI/CD platform with no third-party dependencies, deployed as one binary.
• Lume: An Agent DSL Server in a Single C11 Binary — Why putting business logic in .lume scripts behind one ~1.7 MB binary beats a full runtime stack for small services.

AI Engineering
Agent workflows, tool calling, MCP, RAG, evaluation loops, prompt systems, memory systems, SSE execution logs.

Platform Engineering
Docker, Makefile, multi-stack environments, reproducible workflows, local tooling, developer productivity.

Full-Stack Systems
React, Next.js, TypeScript, Node.js, Python, FastAPI, Redis, GraphQL, REST APIs, Playwright.

FinTech
Customer-facing systems, identity verification, compliance workflows, secure full-stack development.

• PayPal — Senior Full-Stack Engineer (2024 – 2026)
  Customer-facing FinTech systems, identity verification, AI-assisted engineering workflows, full-stack development with React/Next.js/Node.js
• Trip.com Group — Senior Frontend Engineer (2017 – 2024)
  International content platforms, Node.js SSR, SEO, multilingual content, frontend engineering evolution
• Earlier — Software Engineer → Technical Manager (2005 – 2017)
  Enterprise systems, mobile apps, embedded web, cross-team delivery

• Website: https://erishen.cn
• GitHub: https://github.com/erishen
• Email: erishen@qq.com
• Afdian: https://afdian.com/a/erishen