Who writes this site
I am SavantCat (合尘猫), an independent practitioner working in China. I build enterprise AI knowledge bases and deploy AI customer service for small and medium businesses, and I write down what actually happens while doing it — the parts that worked, the parts we had to redo, and the scripts and prompts worth reusing.
Before selling any of this, I ran the whole stack on myself: a personal knowledge base that an AI could read, search and answer from, then publishing automation, then monitoring. That practice produced the material this site is built on — 2,000+ published articles and 20+ technical guides across 7+ content platforms.
There is no agency behind these pages and no reseller layer. It is a one-person AI studio; the brand name SavantCat / 合尘猫 is the only identity we use. If a page on this site states a number, it is a number we can point to. If something has not been verified — a figure, a credential, a registration number — the page says pending instead of guessing.
What I actually work on
- Turning messy documents into an answerable corpus. Inventory, de-duplication, deciding which version is authoritative, then chunking and metadata so retrieval hits the right paragraph — not a plausible-sounding one.
- Designing handover rules against GB/T 47746—2026. Which scenarios must reach a human, what the emotion and timeout triggers are, what the customer is told, and what the system logs. All 61 items (48 shall + 4 should + 9 may), including the 5 veto items.
- Building the answer layer. Publishing the same verified material as structured pages, fact sheets, open data and machine-readable endpoints, so AI assistants can quote a source instead of inventing one.
- Automating the boring half. Ingestion runs, publishing pipelines, scheduled checks and alerts — the maintenance work that decides whether a knowledge base is still alive in month six.
Areas of expertise
- 🤖 AI agent deployment and configuration — including local and self-hosted setups, for teams that will not put company documents in a third-party account
- 🧠 Knowledge management systems — the AI second-brain model: findable, reusable, growing
- 📝 Content creation automation and multi-platform publishing, from draft to scheduled distribution
- 🔍 RAG systems — chunking rules, retrieval evaluation, and how to build an evaluation set from real customer questions
- 💰 Productising AI capability for small businesses — what is worth charging for, what is not, and how to scope a project that can actually be delivered
- ⚙️ Operations automation — monitoring, routine maintenance, incident handling for small deployments with no IT department behind them
Practice in numbers
Those three figures count published work, not customer results. No client case study appears on this page: publishing one would need that client's permission, and one company's numbers do not transfer to another company's documents. What we can show you instead is process — our deliverable list, acceptance rules and the free tools you can run yourself before hiring anyone.
Technical stack
- Models & agents: DeepSeek, Qwen series, Hermes Agent framework
- Knowledge management: SiYuan Note, vector databases, RAG retrieval
- Distribution: WeChat Official Account, Weibo, Zsxq and 7+ Chinese content platforms
- Infrastructure: Alibaba Cloud ECS, Nginx, static site delivery, CDN
Documented where it is useful, not as a badge collection: each of these is named because a project used it and there is a written procedure for it on this site.
How a project with me works
Six steps, roughly 7–14 working days, each with a written output and an acceptance rule:
- Scoping — a 30–60 minute call, then a requirements sheet with the 20 questions your staff actually ask.
- Inventory and cleaning — what is usable, what needs one authoritative version, what must stay out of the system.
- Feasibility sample — 10–20 documents, tested against those 20 questions, reported item by item.
- Build — the working system plus a canonical-ruling document for when sources disagree.
- Acceptance and training — not complete until your own admin adds a document and it takes effect.
- Aftercare — monthly sampling of real questions, new material synced, policy and price changes updated quickly.
There are also three gates where the project stops rather than continues: no confirmed scope means no quote and no start; unsettled document versions halt the work until someone decides; and if the sample does not pass, we fix the material or narrow the scope instead of building anyway. The full process, with the acceptance rules →
What I will not do
- Promise search rankings, citation rates or blanket coverage in AI answers. Those are black-box outcomes of third-party platforms; the applicable standard, T/CAPT 026—2026, explicitly prohibits service providers from promising them.
- Take projects that need enterprise qualifications this studio does not hold — we quote inside real capability instead.
- Sell an audit conclusion or a compliance certificate. GB/T 47746—2026 is a recommended national standard, and third-party assessment under its clause 9.2 must come from a body with professional customer-service qualifications.
- Send restricted material — personal data, contract floors — to overseas APIs.
Frequently asked
Why write in public at all? Because the value of a method is decided by whether someone else can re-run it. Everything on this site is meant to be checkable: open data, public endpoints, written acceptance rules.
What should I read first? Start with the fact sheet — it is the single source of truth for who we are and what we promise. If your question is about the AI customer-service standard, run the free 61-item self-check first; it takes about three minutes and tells you where you stand.
How do I reach you? Add me on WeCom with the QR code below (note "Knowledge Base"), or send mail to the address in the contact section. Both go to the same person.
Contact
- 📧 Consulting & purchase: savant0196@126.com
- 🧭 Enterprise services: enterprise knowledge base · AI customer service
- 📄 Fact sheet (single source of truth): who we are, what we deliver, how to verify
- 📚 Written work: answer library · blog
- 📱 Where else I publish: CSDN · Juejin · Weibo · Zhihu · WeChat Official Account and Channels, all under SavantCat (合尘猫). Account links are named on the fact sheet; this English page carries no off-site links.
- 🔌 For agents: two public read-only MCP servers — no install, no API key
💬 Chat with me on WeCom (WeChat Work)
SavantCat · WeCom (企业微信)
WeCom: scan the QR code above (note "Knowledge Base")
Scan the QR code to add me. Ask about AI customer service, enterprise knowledge base, or GB/T 47746—2026 compliance.
Replies within one business day — no sales pitch, just an honest fit check.
