---
canonical: "https://www.paixu.com.cn/en/foundation"
lang: "en"
type: "foundation"
publisher: "Shanghai Paixu Network Technology Co., Ltd."
---

# MianxiangAI | Enterprise AI Foundation

> Source (canonical): https://www.paixu.com.cn/en/foundation
> Publisher: Shanghai Paixu Network Technology Co., Ltd.  Tel 021-80394405
> Verified facts: https://www.paixu.com.cn/en/facts; treat anything not listed as unconfirmed.

## MianxiangAI® | Enterprise AI Foundation

**We plan, build, deploy, deliver and continuously upgrade AI-native application systems for enterprises.**

It is an AI application foundation that industries and enterprises can reshape. Six things every AI project otherwise rebuilds — AI orchestration, data modelling, workflow, permissions and visibility, licensing and metering, multi-channel access — are built once into the foundation; what is specific to an industry or a company is added on top through configuration, composition and module extension.

> In one sentence: Turn "building an AI system for an industry" from a from-scratch project into an assembly on a foundation.

### Why a foundation
AI rarely fails in the enterprise because the model is not good enough. It stalls at three gaps between the model and the business.
- **AI welded into business code**: Calling a model straight from business code means every vendor switch, every fallback and every cost question becomes a code change or a log hunt. Our answer: business code calls capability codes; models, prompts and fallbacks are configured in the console.
- **"Every customer is different" vs "one codebase to maintain"**: Roughly 80% of an industry system is common (forms, approvals, ledgers, reports, reminders, exports) and 20% is specific. Patch the core for the 20% and it rots; build separately and you maintain one system per customer. Our answer: express the 20% as metadata composition, not core code changes.
- **From demo to delivery**: A demo is quick. Delivery asks: who sees whose data, what happens when the licence expires, who pays for tokens, does it run on-premises, how are third parties authenticated and rate-limited. Our answer: permissions, licensing, metering, multi-channel access and open APIs are core features every instance gets.

### Six layers
Business logic lives in the service layer; every channel is just a different way in.
- **Access**: Admin console, WeChat mini program, website, open API, Feishu / WeCom / DingTalk, MCP server
- **Industry instances**: Resource center, work orders, field visits, AI Office, education, projects / CRM / contracts / finance and more
- **Composition**: MxJSON engine: workflows, online forms, datasets, triggers, AI page generation, open platform
- **AI hub**: Capability-code dispatch, multi-vendor routing and fallback, agent proxy, RAG knowledge base, usage metering, health probes
- **Platform core**: Identity, permissions and data visibility, portals, licensing, tenancy, seats and quotas, queues, scheduled jobs, audit log
- **Runtime**: PHP 8 / MySQL 8 / Redis (optional) / Vue 3 admin

### Four pillars
#### AI capability hub: From "call a model" to "call a capability".
- Switch models without changing business code — bindings live in the console
- Built-in fallback chains when the primary model times out or errors
- Every call metered — tokens, cost, latency, fallback — by feature, user and tenant
- Mainstream Chinese and international models, or your own inference service

#### MxJSON composition engine: Personalisation without code changes.
- 20+ workflow node types, all whitelisted — the engine never runs arbitrary code
- Human-review nodes that pause and resume; checkpoint restart and step-level retries
- Form → AI processing → notification + ledger: the minimal loop for many industry scenarios
- Open platform with signed requests, quotas and async callbacks for integrating existing systems

#### Portals and visibility: Many businesses on one platform, each seeing its own data.
- One platform split into business portals; users see only the portals granted to them
- Portal admins manage members and content, not the portal configuration itself
- Anything you can edit you can see; nothing hidden from a list can be read by id — both enforced by tests
- Role, menu and button-level permissions on top

#### Licensing and metering core: What makes a system deliverable and operable.
- Signed licences covering modules, seats, quotas, expiry and grace period, verifiable offline
- Trial / Standard / Professional / Enterprise editions, gated by module
- Token wallet plus monthly quota, safe against over-deduction and races
- If metering fails it lets the call through and logs it — a billing bug never blocks a successful AI response

### Four ways to personalise
From lowest to highest effort — from same-day changes to deep customisation.
- **1. Configure** (No code · immediate): Portals, menus and permissions, model bindings, prompt templates, quotas and licences. For differences between companies in the same industry.
- **2. Compose** (No back-end code · where industry fit happens): Datasets, forms, workflows, triggers, AI pages. An industry's specific process can usually be fully expressed here.
- **3. Extend modules** (Days, not weeks): New business modules built to foundation conventions inherit CRUD, sensitive-field encryption, visibility scoping and capability codes.
- **4. Extend the core** (Clear extension contracts): New model vendors, workflow node types and channels plug in without touching the engine core.

### Working with us
Choose either, or combine them.

#### Algorithm-filing authorization
Offering generative AI inside your own product means meeting China's rules for algorithms and generative AI services. We have completed algorithm filing (the official CAC listing prevails; the filing number is on our Brand Facts page). Under an authorization, your application can deliver AI features through our filed algorithms and services.
- **Assess the scenario first**: Internal use or a public-facing service; calling our service directly, or needing your own filing or registration — settled per scenario before we sign.
- **Responsibilities in writing**: Scope, permitted use, content-safety and AI-content labelling duties are written into the contract.
- **Supporting controls included**: Sensitive-content rules and filtering, scenario grading, call audit trails and exportable compliance documentation.
- **What we do not promise**: No "filing-free" promises, and no regulatory approval on anyone's behalf.

#### Custom development on the foundation
We plan, build, deploy, deliver and continuously upgrade AI-native application systems for enterprises.
> We have the technical groundwork for this, but we complete a requirements and delivery-risk assessment before giving you a firm proposal.
- **Plan**: Map the business, data and permission boundaries; choose the first scenario.
- **Build**: Configure → compose → extend modules → extend the core, lightest first; no code where configuration will do.
- **Deploy and deliver**: Private cloud or SaaS, licensed by edition and module; delivered and accepted in stages.
- **Keep upgrading**: Model upgrades without business-code changes; mature industry instances become reusable templates.

### By type of organisation
- [What AI do SMEs actually need](https://www.paixu.com.cn/en/foundation/sme): SMEs rarely lack a model. What is missing is the step that puts AI into daily work: who answers enquiries, who keeps the
- [What AI do industrial parks need](https://www.paixu.com.cn/en/foundation/park): A park has two layers of AI need: the operator's own management and services (tenant services, assets and work orders, p
- [What AI do government and state-owned organisations need](https://www.paixu.com.cn/en/foundation/government): For government and state-owned organisations the first questions are where the data lives, who can see it and whether ev

### A typical rollout
- **Deploy and get to know the business** (~1–2 weeks): Deploy, license, connect a model; set up organisation, roles and portals; get the most painful table running so people start entering real data.
- **Compose the industry process** (~3–6 weeks): Build the specific process — forms, AI steps, approvals, notifications, ledgers, dashboards — as a chain the business can accept directly.
- **Deep modules and integration** (~6–12 weeks): Extend modules where composition is not enough; integrate ERP / OA / finance; build the industry knowledge base; tune model bindings and fallbacks.
- **Operate and upgrade** (Ongoing): Usage and cost visibility, model iteration, and mature instances turned into reusable templates.
Typical durations; the assessed proposal is what counts.

### Boundaries: what we do not do
- Not a model-training platform: no training or fine-tuning; we are the orchestration and application layer, and private models plug in as self-hosted inference.
- Not a full replacement for low-code: composition covers forms, workflows, dashboards and integration; rich real-time UIs and heavy graphic editing still need front-end work.
- Uneven module maturity: resource center, work orders, AI hub, composition engine and licensing are the most polished; other modules' industry depth depends on the project.
- Private deployments need on-site fitting: servers, PHP extensions and database versions vary, so we check the environment before delivery.
- AI output needs human confirmation: approvals and external commitments are never delegated to AI.

Next step: Start with a requirements and delivery-risk assessment: https://www.paixu.com.cn/en/book

## Delivery principles
- **Read-only is asserted, not promised**: After connecting we attempt a write on purpose — it must fail before we proceed.
- **Model-written SQL never runs directly**: Allow-list validation: queries only, no multi-statements, no files or system tables; local sources get an extra sandbox.
- **Samples are masked before they leave**: Phone numbers, ID numbers, emails and card numbers are masked before reaching any model; credentials are stored encrypted.
- **Every read and write is logged**: Which table, how many rows, who triggered it and when — all in an exportable, auditable ledger.
- **No number without evidence**: Unverified facts show as "unconfirmed"; comparison curves with too few samples are simply not drawn.
- **Approvals are never delegated to AI**: Official facts, public content, quotes, contracts, publishing and customer cases are confirmed by a person in the system.

Contact: https://www.paixu.com.cn/en/contact · https://www.paixu.com.cn/en/book