Telegram @cangqilai123 · Email · X · All repos
I take AI from a whiteboard idea to a production system a real business runs on — design, build, deploy, operate.
把 AI 从方案设计一路做到生产落地并长期稳定运行：0→1 独立负责，交付即可用。
qilai / tytsxai · AI Forward Deployed Engineer (FDE) · China · GMT+8 · open source under tytsxai
Keywords: AI FDE · LLM engineering · multi-agent · RAG · production AI · self-hosted bots · automation · Telegram tooling
POSITION
Looking for AI Forward Deployed Engineer (FDE), AI application engineering, LLM / agent engineering, or intelligent automation. Onsite in China or fully remote, worldwide.
Who this profile is for / 适合谁： hiring managers and founders evaluating an FDE who ships production AI systems end-to-end — not demos only.
- Born in the 2000s. Four years of hands-on AI engineering — shipped systems, not slideware.
- I own the whole line: solution design → build → deploy → operate. Several production systems still running.
- Depth in agents, multi-agent orchestration, RAG, automation matrices, evals, observability, cost control.
- Frontier tech with a cost ceiling. Token and infra spend is a requirement I design against, not an afterthought.
- Results-oriented, and honest about two-way fit. Send me the problem rather than the job description.
中文版
00 后，目前在国内，寻找国内或远程的 AI Forward Deployed Engineer（FDE）、AI 应用落地、LLM 开发、多 Agent 系统、智能自动化岗位。护照正常，可出差、可出境。
4 年 AI 实战经验，擅长从 0 到 1 独立负责，从方案设计一直做到生产落地，已交付多个长期稳定运行的项目。熟悉 Agent 编排、多 Agent 协作、RAG 检索增强、自动化矩阵等方向。注重前沿技术与成本控制的平衡，能高效把 AI 转化为可衡量的实际价值。
结果导向，希望加入真正做事的结果导向团队，双向匹配。欢迎直接私信：Telegram @cangqilai123。
THE ROLE
An FDE works in the gap between a model that demos well and a customer whose reality is messy. Four years of closing that gap:
| The job really is | What that looks like in my work |
|---|---|
| Sit with the customer, find the real workflow | I start from the operator's critical path and failure modes, not from the model |
| Build a working system in days, not quarters | 0→1 solo delivery: design → MVP → docs → CI → production |
| Survive contact with production | Dry runs, review gates, rate limits, retries, rollback paths, structured logs |
| Keep unit economics sane | Model routing, caching, batching, context budgeting, cheap-model-first tiering |
| Hand it over so it keeps running | Bilingual runbooks, deploy scripts, systemd and CI, redaction gates |
| Cross language and culture | Native Chinese, working English, bilingual docs as a default habit |
EVIDENCE
Public, runnable, maintained. Stars are users, not followers.
| Project | What it is | What it proves |
|---|---|---|
| IDM-Activation-Script-Chinese | Windows batch toolkit · GPL-3.0 · ~1k stars · 75 forks | Distribution and maintenance at real scale — issue load, GBK/CP936 encoding traps, registry backup, environment self-check |
| PromptPanel | macOS prompt and snippet launcher for AI power users · Swift/SwiftUI | Product thinking applied to the LLM workflow itself · local-first UX · release-readiness CI |
| bilibili-cleaner | Bulk account cleanup · FastAPI + Web UI + Typer CLI + Docker | Destructive automation done safely: QR login, rate limits, review gates, pytest CI |
| macfriends-cli | WeChat relationship detection on macOS · Rust + ObjC++ agent | Systems depth: native agent, ABI-pinned reverse work, strictly local execution |
| anyreality-resi-stack | Self-hosted sing-box infra: installer, subscription server, dual-node routing | Infra and ops delivery: bilingual runbook, systemd units, secret-redaction CI |
| bazi-master | Full-stack AI divination platform · React + Express + PostgreSQL + Prisma | LLM inside a product: interpretation pipelines, multi-module domain logic, full-stack ownership |
| x-account-cleaner | Local-first X/Twitter cleanup CLI · TypeScript + Playwright | Browser automation with a review-first, irreversible-action-aware model |
| telegram-ui-builder · demo | Visual designer for bot messages, inline keyboards, multi-screen flows | Developer-facing tooling with a deployed demo and Pages CI |
A meaningful share of my work is private client and production systems — agent pipelines, ops automation, data workflows. Happy to walk through the architecture and the trade-offs in a call.
TOOLS
METHOD
Start from the workflow, not the model. Who runs it, what breaks, and what "done" means in their numbers.
Smallest useful loop first. A runnable thin slice in days, then harden against the friction that actually shows up.
Treat risky actions as first-class. Dry run, confirm, limit, log, roll back. Visible failure beats silent failure.
Measure before claiming. Evals, reproduction steps, real runs. "Verified" and "should work" are different words.
Package for the next person. Docs, deploy scripts, CI, and runbooks, so it outlives my involvement.
WORK
Every visual here is generated from the GitHub API by scripts/gen_assets.py and refreshed weekly by CI. Self-hosted on purpose: third-party card services go rate-limited and render as broken images, and a profile that breaks is a profile that says the wrong thing.
CONTACT
Telegram @cangqilai123 — fastest. Also email, X, V2EX.
Send the problem, not the JD. Within a message or two I'll tell you whether I can actually solve it and roughly how.
最快联系方式：Telegram @cangqilai123。直接说要解决什么问题就行，我会告诉你能不能做、大概怎么做。
AI is only worth something once it survives production. Automation is only worth something while it stays under control.