chenk.top
中文
Portfolio · 2026

Long-running agents — and the
system foundations that sustain them.

Marketing automation. Autonomous research. Agent control planes. Three distinct threads — bound together by the quiet, critical infrastructure: observability, token economics, skill evolution, shared memory.

9 projects Typically replies within 24h
Filter
public 2026
Trends Copy · Image · Video · TTS Atomize · Debate · Storyboard Templates · Calendar · i18n Multi-publish · Analytics

AI4Marketing

Product · Full-stack · AI orchestration

One sentence in, full campaign out. Trend discovery → AI copywriting → image generation → short-video production (narration / drama / talking-head) → TTS → 9-language localization — fully automated. Spark Drama Studio (agentic screenwriting → directing → review loop), GEO optimization (make AI search engines cite your content), content atomization, multi-perspective debate, storyboard editing, content calendar, and an admin dashboard. 121 API routes, 21 database tables, 4 payment channels, built on the full DashScope AI model suite.

Next.js 14 TypeScript DashScope Qwen3 HappyHorse 1.0 PostgreSQL Prisma CosyVoice TTS Aliyun OSS
121API routes
37pages
9languages
llm4marketing.com
public 2025–now

DaaS — Documentation as a Skill

Skill pipeline · GEO platform · Agent toolchain

Distill a company's docs into agent-ready Skills.

Point it at a folder of product docs. An LLM reads everything and writes two kinds of Skill: detail skills (how to use each capability) and intent skills (which path to pick when several lead to the same goal). Output plugs directly into Claude Code, Cursor, or any agent that reads SKILL.md. Zero human authoring, zero third-party runtime dependencies (pure Python stdlib). GEO closed loop (measure → diagnose → rewrite → re-measure) ensures products get cited by AI search engines; edge-exploration agent auto-discovers cross-product combo scenarios. Ships with auto-generated MCP servers, D3 knowledge graph, live SSE telemetry, and dual-currency subscription billing.

zero-dep Python DashScope MCP D3 SSE GEO flywheel
478+auto-gen Skills
17products
96%adversarial recall
company-skill.com
public 2026

MiniGameForge

AI game workshop · elevator runtime

Pick a template, pick a style, hit generate. AI auto-produces game code, art assets, and cutscene videos. Powered by the elevator runtime (planner → executor → verifier), 7-category tripwire guardrails, and visual self-debug — adaptive routing across 6 providers and 17 API keys.

elevator visual self-debug tripwires Qwen
llm4marketing.asia
public 2024–now

chenk.top

Tech blog · Bilingual EN/ZH · Hugo

Both Chinese and English written from scratch — never translations. Chinese leans concise, English leans expository. 30+ thematic series spanning ML, mathematics, cloud computing, and agent systems, with a custom Hugo theme featuring LaTeX rendering, full-text search, series navigation, and dark mode.

Hugo KaTeX bilingual custom theme
692articles
33series
www.chenk.top
semi-public 2025–now

AI4Science

Autonomous research · Knowledge graph · Distributed agents

A fully autonomous research system that runs 24/7.

Two independent pipelines work in tandem: the knowledge accumulation pipeline discovers papers, performs deep reading, maintains a knowledge graph (65K+ nodes, 197K+ edges), and generates research ideas through adversarial debate. The execution pipeline, driven by a Boss Agent (LLM-orchestrated), takes approved ideas through experiment design, execution, statistical analysis, failure triage, manuscript writing, and peer review. Distributed fleet (2 worker nodes, 128GB RAM each), self-healing supervisor (40+ rules), and explicit FSMs ensure full auditability.

Public showcase at llm4science.top exposes the research feed; orchestration logic and private datasets stay internal.

Knowledge Graph adversarial debate Boss Agent distributed fleet self-healing
41K+papers read
65K+KG nodes
4K+ideas generated
1.1K+manuscripts
llm4science.top (semi-public · partial results shown · full system in backend)
internal 2026

llm-elevator

Autonomous coding · Multi-model orchestration · DAG execution

Enable Chinese LLMs to autonomously complete complex software engineering tasks.

A high-level goal decomposes into a DAG of verifiable subgoals. Different models handle planning, execution, review, and verification — no model grades its own work. Failures auto-escalate through three model tiers (Qwen → DeepSeek → strongest available), subgoals execute in parallel, tripwires detect 7+ behavioral anomaly patterns in real-time, and an experience system compresses failures into reusable skill templates. Supports multi-day projects (Roadmap → Milestone → Task → Subgoal DAG) with continuous autonomous advancement.

Qwen DeepSeek Kimi DAG execution cross-model verify skill evolution
5model families
3-tier escalation
7+tripwire patterns
Internal development system, not publicly available
public 2026

OpenClaw QuickStart

Tutorial repo · Self-hosted agent

OpenClaw is a self-hosted gateway that bridges 20-plus chat platforms into one agent runtime. This QuickStart breaks deployment, model integration, channels, skills, and cron into 10 bite-sized, hands-on guides — each takes <15 mins to complete.

openclaw self-hosted skills mcp dingtalk telegram
github.com/chenkai66/OpenClaw-QuickStart series · 10 parts
public 2026

Claude Code Hands-On

Tutorial repo · Programmable dev platform

Most people use 10% of what Claude Code can do. This hands-on guide unlocks the other 90% — three-layer config, thinking modes, slash commands, MCP, hooks, the SDK, and the GitHub Action.

claude-code mcp hooks sdk github-actions
github.com/chenkai66/claude_code_learn series · 10 parts
Currently exploring

A few things on my desk

01

Agent token economics

How should a long-running agent split its budget across providers? Which steps are worth the most expensive model?

02

Limits of skill evolution

Which failures can compress into reusable skills, and which can only be discarded?

03

Schemas for shared memory

When multiple agents share one memory, how strict should the type system be?

04

Prompt to product

From demo to production — the final mile: — observability, replay, governance, and user trust.

Get in touch

Interesting projects, collaborations, or just ideas — all welcome.

chenkai.nb.666@gmail.com GitHub
writing 2026

LLM App Security

Security engineering · Methodology · Open-source book

A practitioner's guide to securing web applications in the AI Coding era.

When AI writes 80% of your business logic, where do the security boundaries live? This book dissects the attack surface of LLM-powered applications — prompt injection, privilege escalation, data exfiltration, supply-chain poisoning — and delivers engineering-grade defense-in-depth: pre-commit secret guards, TOCTOU quota-race protection, multi-layer firewall coordination, agent behavior sandboxing. Every pattern is extracted from real production incidents.

Security LLM Defense-in-Depth Open Book
github.com/chenkai66/llm-app-security