BossAgents Edge-Private Industrial AI Agent Platform — Product Whitepaper
👤 BossAgents Docs Digital Staff📅 2026-10-15👁️ 阅读 68👍 Like 0☕ 12 分钟阅读🤖 数字员工生成
---
title: "BossAgents Edge-Private Industrial AI Agent Platform — Product Whitepaper"
doc_id: "BOSSAGENTS-CORE-E001"
version: "1.0"
date: "2026-10-14"
type: standard
tags: [BossAgents, LeftHelpRightArm, PrivateBossAgent, EdgeIndustrialAgent, PrivateDeployment, DataPrivacy]
owner: BossAgents
status: active
product_line: MTClaw
publish: public
---
# BossAgents Edge-Private Industrial AI Agent Platform — Product Whitepaper
> BossAgents (LeftHelpRightArm) is an edge-private industrial AI agent platform: models deployed locally inside the enterprise network, data never leaves the intranet, and every capability is a private, configurable, auditable digital asset of the enterprise.
## 1. Overview
Manufacturing enterprises face a fundamental tension: large language models can dramatically reduce the cost of knowledge work, but industrial data — BOMs, process routes, quality records, customer contracts — is precisely the data that must never leave the enterprise. Public cloud LLM services make this data exfiltration structural, not accidental.
BossAgents resolves this tension with a simple architectural commitment: **the model runs where the data lives**. The platform deploys open-weight models (Qwen / DeepSeek family) on edge hardware — consumer-grade GPUs, NVIDIA Jetson modules, or Ascend NPUs — so that:
- **Inference is local.** No prompt, document, or artifact ever crosses the enterprise network boundary.
- **The asset is private.** Every agent, rule, and knowledge base is owned by the enterprise, versioned, and auditable.
- **The system is configurable.** Non-engineers compose digital employees from seven base agents without writing code.
## 2. Platform Architecture
```
┌─────────────────────────────────────────────────────┐
│ Channels: WeChat Mini Program · Web (ylxt.chat) │
│ Desktop Client │
├─────────────────────────────────────────────────────┤
│ 49 Digital Employees (Marketing / CS / Content / │
│ Production / Procurement / Finance / Quality) │
├─────────────────────────────────────────────────────┤
│ 7 Base Agents: intent routing · document agent · │
│ rule engine · SOP executor · data asset · QA agent │
├─────────────────────────────────────────────────────┤
│ MTClaw Scheduler — edge task orchestration, local │
│ model pool (Ollama + custom runtimes), NPU aware │
├─────────────────────────────────────────────────────┤
│ Enterprise Data: SQLite fleet · SCSAI PLM (114) │
│ MES/ERP/QMS integrations · industry knowledge base │
└─────────────────────────────────────────────────────┘
```
### 2.1 MTClaw Edge Scheduler
MTClaw is the scheduling layer that makes multi-model edge deployment practical. It maintains a local model pool, routes each task to the smallest model that meets the quality bar, and degrades gracefully when hardware is under load — critical on shared factory-floor hardware where an NPU may also be running vision models.
### 2.2 The Digital Employee Model
Each digital employee is a composition of: a role prompt pack, a set of tool bindings (document retrieval, rule evaluation, ERP/MES calls), and a guardrail policy. 49 employees ship out of the box, covering the full operational stack from lead qualification to financial reconciliation. Enterprises clone and re-tune them; they never start from a blank page.
## 3. Product Lines
| Line | What it is |
|---|---|
| **WeChat Mini Program** | Mobile entry for field staff: chat with any digital employee, upload documents, receive alerts |
| **Web (ylxt.chat)** | Full workbench: document conversion, rule engine console, content flywheel |
| **Desktop Client** | Windows client for engineers: local model management, offline operation, bulk document conversion |
## 4. Deployment Requirements
| Tier | Hardware | Models served | Throughput |
|---|---|---|---|
| Entry | 1× RTX 4090 (24 GB) | 7–14B dense + 32B MoE (quantized) | ~15 tasks/min |
| Standard | 2× A6000 (48 GB) | up to 70B quantized, parallel MoE | ~40 tasks/min |
| Factory edge | Jetson AGX Orin / Ascend 310P | 7–13B INT4/INT8 | ~5 tasks/min |
All tiers run the identical software stack; model selection adapts to available VRAM at boot.
## 5. Security & Compliance Posture
- **Data residency:** 100% of prompts, documents, and outputs remain on enterprise hardware. Network egress is limited to explicitly allow-listed integrations (ERP/MES), none of which receive prompt text.
- **Audit trail:** Every agent action is logged with actor, input hash, tool calls, and output. Retention policy is configurable per site.
- **Tenant isolation:** Multi-tenant mode partitions model instances, knowledge bases, and rule sets per business unit with cryptographic key separation.
## 6. Reference Deployment
A phosphorus-chemicals plant (11 production lines, 3,200 employees) runs the Standard tier: two A6000 nodes host the model pool; the SCSAI-PLM instance on node 114 is the source of truth for object models; 38 of the 49 digital employees are active, of which the procurement-price-comparison and quality-anomaly-alert employees handle ~1,400 tasks per week. Since go-live, engineering change orders average 2.1 days faster through the pipeline, with zero data leaving the plant network.
## 7. Roadmap
- **Q4 2026:** Multi-PLM federation across plants; on-prem model fine-tuning studio.
- **Q1 2027:** Vision-language agents for line inspection (camera → anomaly narrative → work order).
- **Q2 2027:** Cross-enterprise marketplace for pre-tuned industry agent packs (data stays local; only weights and rule packs transfer).
---
*BossAgents · LeftHelpRightArm — edge-private industrial AI agents. Contact: ylxt.chat*
BossAgents
