// AI Full-Stack Engineer · Hangzhou, China · Remote-friendly

I build AI agent platforms and the full-stack products around them.

Nine years shipping production software. Currently engineering an AI-for-Science agent platform — LangGraph runtimes, RAG with citation tracing, workflow engines — and before that, frontend infrastructure used by tens of millions of users.

● Open to freelance / part-time ● Open to full-time remote UTC+8 · overlaps EU mornings & US evenings
View live demos ↓ maxinohnono@gmail.com
9 yrsengineering production systems
170+merged MRs on current agent platform
21public repos · npm-published SDK
0LLM API keys needed to try my demos
Curry Ma
Curry Ma · 马鑫AI platform eng by day, infrastructure tinkerer always
Featured work — everything below runs live

Demos you can actually click

Self-Hosted RAG Playground

BUILT FROM SCRATCH · SELF-HOSTED

A complete retrieval-augmented-generation pipeline running entirely on my own 2-core VPS — no external API, no cloud vector database, nothing to sign up for.

  • Upload a document (PDF / Markdown / TXT) and watch it get chunked, embedded and indexed in real time
  • Hybrid retrieval — SQLite FTS5 keyword search fused with local ONNX vector embeddings (Reciprocal Rank Fusion)
  • Every answer snippet carries a citation back to the exact source chunk — the anti-hallucination pattern I build at work
  • Optional "Ask" mode is bring-your-own-key: your OpenAI-compatible API key is used in-request only and never stored server-side
Python · FastAPISQLite FTS5ONNX embeddings (bge-small)RRF hybrid searchVanilla JS frontendDocker + Traefik
Open live demo ↗ Source soon ↗

Turbo Support Platform

FULL-STACK SAAS

Dual-app customer support & ticketing SaaS — separate admin and customer applications sharing one Turborepo monorepo.

  • Real-time chat over WebSocket with an AI assistant layer and human-agent takeover
  • Tenant & role data isolation enforced at the database level via Postgres row-level security
  • Shared packages: database client, auth, UI kit, support core
Next.js 14TurborepoSupabase RLSStream ChatTypeScriptshadcn/ui
User app demo ↗ Admin demo ↗ GitHub ↗
admin demo login: superadmin@mx / 123456 · user app: register your own

omni-log

NPM-PUBLISHED SDK

Zero-instrumentation logging & monitoring SDK for WeChat / Alipay mini-programs and web. Install it, and it quietly captures everything.

  • Auto-captures page lifecycle, HTTP requests (params, duration, timeouts), JS errors with full stacks, unhandled rejections, clicks
  • Web build adds Core Web Vitals: FCP / LCP / CLS / INP / TTFB
  • TypeScript monorepo: log-core / log-web / log-mini / error packages
TypeScriptnpmMini-programWeb VitalsMonorepo
npm package ↗ Live demo ↗ GitHub ↗
More on GitHub: rubick-plugin-monalisa · design-diff automation (puppeteer + blink-diff) api-previewer · lightweight API response inspector · 21 public repos → github.com/mx94
Current role

What I'm building right now

Molpe AI — AI-for-Science agent platform

2026.05 – present · Shanghai / product: MoLex ↗

A platform where scientists run protein-structure analysis and molecular design through AI agents — long-running reasoning tasks, tool calls into scientific models, human-in-the-loop approval.

Agent runtime: self-built LangGraph-compatible gateway (Django ASGI) with durable runs, checkpoint/resume, Redis-Streams streaming and reconnect replay
RAG knowledge base: ingestion → chunking → hybrid retrieval → rerank, with answers traceable to source files via citation linking
Molecular design workbench: interactive Mol* 3D viewer with a selection-first UX and AI design copilot, wired to FoldX free-energy evaluation
Workflow engine: drag-and-drop DAG canvas, graph validation, resumable execution, MinIO artifact storage
Agent observability: observation/event model — per-step tool I/O, parent-child spans, deterministic problem signals
Platform quality: schema contract tests across 13 models, connection-pool governance, CI parallelization & release discipline
React 19 + TypeScriptLangGraphPython 3.12 · Django · CeleryPostgreSQL · Redis · MinIOMol* / NGLmypy strict
Track record

Experience

2026.05 – presentShanghai, China

Molpe AI

AI Full-Stack Engineer
  • Core engineer on the bio-agent platform: agent runtime, RAG pipeline, molecular design workbench, workflow engine (details above)
  • 170+ merged MRs; peak week 157 commits — platform engineering plus product features end to end
2022.09 – 2026.03Hangzhou, China

Qingtuan — flexible-staffing platform

Senior Frontend Engineer
  • Built the mini-program error-tracking SDK used across the fleet — auto-alerts and auto-created tickets caught most production errors before users reported them (later open-sourced as omni-log)
  • Shipped a streaming-output Feishu bot that queried SLS logs in natural language — early LLM tooling before it was trendy
  • Created a macOS SwiftUI build tool that moved mini-program CI/CD off the overloaded central build server onto every dev's machine (10+ framework/Node combos)
  • Config-driven component system (search filters / tables) adopted across business teams, with screenshot-to-admin-code generation
2020.09 – 2022.09Hangzhou, China

Zhiyun Health — chronic-disease management (HKEX-listed)

H5 Frontend Engineer
  • Owned "Karl" — a low-code page-builder platform that shipped 287 production H5 campaign pages, cutting launch time from 1–2 days to ~2 hours (Vue workbench + Node server + GitLab CI/CD)
  • Built the Rubick Electron packet-capture/mock tool and the Monalisa design-diff automation — 10+ project teams adopted them, UI-fidelity scores up 20%
  • Delivered the patient-side e-commerce flow and WebSocket-based health-coach IM used across hospital SaaS products
2017.07 – 2020.09Nanjing, China

Kidswant — mother-and-baby retail (ChiNext-listed)

R&D Engineer
  • Mini-program live commerce: portal + live-room purchase flow — cut live-room open wait by ~3s, smoothed likes/danmaku realtime rendering
  • Seckill & coupon marketing tools running inside dual containers (App webview + mini-program webview) with runtime bridge detection
  • Server-side PHP components for marketing campaigns — coupon issuance/redemption, lottery with inventory control, async push/SMS
Toolbox

Skills

AI & Agents

LangGraphdeepagentsRAG — hybrid retrieval · rerank · citationsMCPHITL workflowsContext engineeringStreaming (SSE · WebSocket · Redis Streams)ONNX local inference

Frontend

React 19TypeScriptVue 3Next.jsWeChat / Alipay Mini-ProgramsMol* · NGL (WebGL)@xyflow canvasTailwind / shadcn

Backend & Data

Python · Django · FastAPIPostgreSQLSQLite FTS5RedisCeleryMinIONode.js (egg.js)Supabase RLS

Platform & Quality

Docker ComposeTraefik · Let's EncryptGitLab CI/CDContract testingmypy strictObservabilityThis very site: self-hosted on a $5 VPS
Work with me

Need an AI feature shipped, end to end?

I take on remote freelance and part-time engagements: agent / RAG / chat product development, full-stack builds (Next.js + Python), and performance rescue on existing frontends. Typically 10–15 focused hours a week, replies within 24 hours, starting with a concrete technical plan — not a sales call.

Email me — maxinohnono@gmail.com Browse GitHub ↗