TIO AUTOPLAT · Intelligence layer · AIHub V1.6

AIHubEnterprise AI & knowledge foundation

AIHub turns capabilities scattered across cloud models, agent platforms, business systems, documents and memory into shared enterprise intelligence. On the AI side it connects mainstream LLMs, industrial small models, external agents and business tools, orchestrated by a cognitive-agent runtime; on the knowledge side a Wiki structural hub plus RAG and memory organizes enterprise knowledge. Digital employees are built on both.

75%+

autonomous decision rate

85%+

tool-call accuracy

95%+

resume-after-interrupt rate

99.99%

traceable executions

AIHub

Orchestration

Topology DAG, dynamic dispatch, HITL, sub-agents

Context memory

Memory tree, trust governance, equipment and customer history

Knowledge retrieval

Wiki L1/L2 routes first, Fusion RAG fetches details

Closed-loop execution

MES, small models, WeChat/phone/email, MCP / Skill

Execution evidence flows back to memory and Wiki

* Design targets; final figures depend on the released version and project evaluation.

Enterprises do not need another chat model; they need an AI and knowledge foundation

The real gap is that models, knowledge, business semantics and actions are still scattered across people, systems and channels.

Before

People at the center

· Read SOPs and process documents
· Log in to MES / ERP / CRM
· Chase by phone, coordinate on WeChat, align by email

Work is stitched together by people; experience is hard to copy and hard to trace.

First wave

Chat on a screen

· Answers questions
· Forgets equipment history
· Can't read the production structure or finish cross-system work

The interface changed; the way work is done did not.

Now

Role-level work systems

· Carries enterprise structure and role memory
· Organizes multi-step collaboration
· Changes systems and reaches people within a whitelist, feeding evidence back

From chatting to getting work done.

Two foundations: AI and knowledge

The AI foundation answers which capabilities to use and how to finish the work; the knowledge foundation answers which enterprise path to follow. Without knowledge, paths drift; without an AI runtime, a graph is just a static archive.

AI foundation

· Models: mainstream LLMs plus industrial vision, acoustic and time-series small models
· Capabilities: external agent platforms and business systems such as MES / ERP connected through MCP tools
· Cognitive-agent runtime: orchestration, memory and execution, human review and re-planning
· Open and governed: OpenKey + MCP + Skill, capability allowlists, tenant quotas and audit

Knowledge foundation

· Wiki structural hub: how the business is organized and which path to take now
· RAG for details: L1 routing first, then drill into documents and the graph
· Memory: situational memory of equipment, shifts and roles, with trust and correction
· Compilation: execution evidence drives local compilation, merged after proposal approval

A digital employee = a role configured on both foundations

Not a model account, but a role unit with structured knowledge, memory and execution boundaries. To copy a role, swap the Wiki binding and tool whitelist — no need to rewrite its persona.

Digital employee =

Role goal

What this kind of work must achieve

Agent topology

Steps, branches and human review (HITL)

Wiki binding

Which enterprise path to follow right now

Memory policy

What is already known about this machine or supplier

Tool / Skill whitelist

Which systems it may change and which channels it may use

Copyable

Roles copy across plants, lines and collaboration scenarios

Auditable

Every basis, decision, tool call and system change is recorded

Improvable

Successes and failures become memory, evidence and structure proposals

Governable

Wiki proposals and tool whitelists control what goes live

How a digital employee handles a line incident end to end

Every step happens in the same orchestration context, with memory loaded and a path already matched.

AlarmOrchestrate

A telemetry event triggers the role topology

Load memoryMemory

Recalls 90 days of alarms and maintenance for this machine

Match pathKnowledge

Wiki L1 finds the equipment-family policy and spare-parts BOM

Human reviewHITL

Stop-line and changeover decisions confirmed via HITL

ExecuteExecute

Freeze the MES batch → notify the team on WeChat → call the parts supplier → email the customer

Feed backFeedback

Root cause and handling time go to memory; gaps become Wiki proposals

What changes is not the interface, but who organizes the work

People move from driving the process to approving key steps; digital employees carry memory and paths across systems.

Work	Traditional	Digital employee
Line incident	Shift lead reads the SOP, then instructs the floor	Loads equipment and shift memory + Wiki hits the process path + orchestrated advice, writes to MES when needed
Anomaly diagnosis	Engineer checks history and judges manually	Memory surfaces past alarms + industrial small model + evidence fed back
Cross-system scheduling	Planner switches between ERP/MES/WMS	Task triggers topology + HITL approves key changes + full trace
Process Q&A	New hires ask veterans and read paper SOPs	Wiki L1 locates the process path + RAG drills down, read-only
Quality inspection	Visual check, manual logging, lead review	Vision model screens + branch review + re-inspect or auto-pass + evidence fed back
Multi-party coordination	Back-and-forth by phone, WeChat and email	System actions and channels dispatched together under one trace

Wiki answers the path, RAG answers the details

Wiki is not a better knowledge base but the structural hub and runtime routing layer for enterprise knowledge. Every finished job makes the structure more stable.

Sources

Libraries, session summaries, high-confidence memory, DataHub catalog contracts

→

Local compile

Incremental, one source work item at a time

→

Proposal gate

No evidence, no commit; merged after approval

→

Runtime routing

Healthy graph routes via Wiki, with fallback

→

Feedback

Evidence and failures flow back into compilation

The L2 property graph holds entities, relations and evidence, with health levels and drill-down navigation.

	Traditional RAG	AIHub Wiki
How it finds	Full-text/vector hits on fragments	An L1 decision map guides dispatch, then drills into L2 and documents
What you see	A handful of fragments	Decision paths, relations and structural health on the graph
How it improves	Load more documents	Execution evidence drives local compilation, merged after approval

Real console screens

From role design and acceptance testing to runtime audit, every step has a visual workbench.

Steps, branches, human review, dispatch and execution become nodes.

Five capabilities that put employees to work

Unified work

One employee unit per kind of work, fixed in a topology instead of fresh prompts each time.

Routable knowledge

Wiki paths plus RAG details, consumed by dispatch in one context.

Engineered orchestration

HITL interrupts, resumable tasks and replanning on failure.

Auditable execution

From reasoning to tool calls and system changes, every step is recorded.

Open and extensible

OpenKey + MCP + Skill with capability whitelists, tenant quotas and scope isolation.

75%+

autonomous decisions

15+

turns of background memory

85%+

tool-call accuracy

95%+

resume rate

80%+

replanning success

≤5%

hallucination rate

99.99%

traceability

95%+

long-chain explainability

* Design targets; final figures depend on the released version and project evaluation.

The difference is not the model brand but whether AI works with enterprise knowledge

Successes enter the structure, failures become proposals: a foundation, not a chat wrapper.

Dimension	Typical agent platform	AIHub enterprise AI & knowledge foundation
Unit of work	One chat / one prompt	A configurable, copyable digital employee per kind of work
Memory	Piled-up chat history	Five-layer memory tree with trust, toxicity and correction, in the execution context
Knowledge	Single vector RAG	Vectors + Wiki L1 routing + L2 graph, consumed by dispatch
Execution	A few built-in tools or plug-ins	Dispatches MES, small models, WeChat, phone and email
Models	One LLM	LLM orchestration + industrial vision, acoustic and time-series models
Orchestration	Linear agent or simple chain	Visual topology DAG + dynamic dispatch + sub-agents
Recovery	Retry or fail	Resume, replan, and turn failures into proposals
Flywheel	None	Evidence flows back to memory and Wiki; it gets steadier with use

Four frontline scenarios

Structured knowledge is always at work; write access is granted by the role's tool whitelist.

MES changeover / stop-line

Alarm-triggered with key parameters under HITL; equipment and shift memory; Wiki hits the changeover process; MES orders and batch freeze plus multi-channel notices.

Industrial inspection

Pass, re-inspect or escalate by confidence; vision model plus MES pass-through; defect evidence back to the library and Wiki.

Andon incident handling

Remote reset, on-site repair or parts swap; 90 days of alarms and maintenance recalled; CMMS/MES plus multi-channel coordination.

Read-only shop-floor assistant

Read-only topology that drills along routings and SOPs, returning paths and citations; misses become knowledge-gap proposals.

In AUTOPLAT

AIDA is the private agent people use every day; AIHub supplies AI and knowledge; DataHub supplies governed data; AIBase trains models.

Build an AI and knowledge foundation for the enterprise

Book a demo and see models, knowledge and agents run on one foundation, with a kind of work configured as a digital employee.