From demo to production.
I'm Liu ZhuoQi, an AI Application Engineer focused on Agent systems. I mainly build AI applications, Agent workflows, and production delivery systems.
I served as technical lead and led development for AISEO, Help Center, and UMS. I have also deployed and integrated SGLang / vLLM services for a video-understanding Agent. I now work on a model API aggregation platform at an AI infrastructure company: integrating and load-balancing many upstream model providers, and owning billing for video and image models, onboarding verification for new providers, and upstream quality monitoring.
BUILD
AI Applications & Agents
Multi-stage Agent Workflows · RAG · Tool Use · Multi-Model Routing · Feedback Loops · Agent Evaluation
Frontend
React · Vue 3 · TypeScript · Vite
SERVE
Model Serving
SGLang · vLLM · Multimodal Video Understanding · GPU Integration and Benchmarking
Backend & Workflows
Python / FastAPI · Java / Quarkus · Go / Gin · Temporal · PostgreSQL · Redis
SHIP
Engineering Delivery
Linux · Docker · Kubernetes · Jenkins · GitHub Actions · Ansible · Prometheus / Grafana
$ ls projects/
Projects & Responsibilities
Model API Aggregation Platform
Engineer · Current roleMulti-upstream routing and load balancing, per-second / per-resolution billing for media models, and provider consistency verification.
Vane
Independent · Paused (2026-07 to 09)Built from scratch to production with Go / PostgreSQL / Temporal and React / TypeScript: workflows, a Feishu Agent, and a profile feedback loop.
proj_03Argus
Open source · Long-video understandingA frontend-only long-video understanding agent harness: video never leaves the browser, multi-provider LLMs, frame / memory / sub-agent tools.
AISEO
Technical Lead · AI ApplicationBuilt a multi-stage Agent content pipeline with Python / FastAPI / Temporal, covering RAG, multi-model routing, token costs, and multi-CMS publishing.
Help Center
Technical Lead · Knowledge PlatformDeveloped a multi-site knowledge platform with Java / Quarkus / React, integrating RAG retrieval, vector storage, and local embeddings.
UMS
Technical Lead · Unified PlatformBuilt a multi-product management platform with Go / Gin / Casdoor for shared authentication, permissions, subscription plans, and entitlement sync.
My proficiency varies across technologies. I care more about owning development, testing, troubleshooting, and deployment outcomes than presenting every framework I have used as an area of mastery.
$ ls posts/ --pinned
Technical Writing
- Check the Goods Before You Buy a Relay API
Response id prefixes, input tokens that match the official counting endpoint one for one, the error a deliberately invalid parameter gets, the C2PA signature inside a video: how to check each and what it costs. The whole procedure ships as an open-source Agent skill, llm-relay-audit, that installs in one line and writes the comparison report.
- Agents That Check Their Own Work and Tidy Their Own Memory
Outcomes and Dreaming taken apart from the official docs, Claude Code's /goal tested 12 times on one task, and the design turned into two contracts you can carry to any Agent stack.
- Five Codex Harness Designs Worth Copying
From the openai/codex source: selective parallelism, Unified Exec, ExecCell, progressive Skills, Goal continuation, and embedding the same runtime through App Server.
- Why Codex Does Not Give the Model Every Tool
From the rust-v0.147.0 source: why tools load on demand, how BM25 ranks the best few out of many, and how to carry the same design to Python, Go, and other models.
Say hi
Tinkering with AI apps, Agent engineering, or the infrastructure underneath them too? Any channel above works — just say hi.