Ouxiaojiu

An AI agent for R&D compliance management

Ouxiaojiu checks and analyzes the company’s project, hours and expense data against the policy basis, drafts initiation plans and runs compliance checks and panoramic diagnosis. Every finding cites its sources. Operations that change data require the user’s own confirmation or approval.

14
AI capabilities
200
atomic diagnosis checks
52
MCP tools
3
permission levels

Agent matrix

14 AI capabilities, managed by permission level

Each AI capability has a defined scenario, the data it can access and its operation permission. Query capabilities do not change business data, and operations that change data take effect only after approval or the user’s own confirmation.

  • Read-only
  • Approval required
  • Human confirmed

AI agentsCall business tools on their own to complete multi-step analysis and processing

R&D management assistant

Read-only

Answers R&D management and policy questions using the company’s own project and expense data.

AI panoramic diagnosis

Read-only

Runs 200 atomic checks on R&D projects across five dimensions and produces a diagnosis report and remediation plan.

Project compliance check

Read-only

Checks a project’s initiation, hours, expenses and materials item by item to identify non-compliance early.

Policy compliance Q&A

Read-only

Answers with reference to policy texts and company data, and every answer can be traced to its sources.

AI initiation planning

Approval required

Generates the initiation plan section by section from the R&D direction, shortening the time needed to prepare initiation materials.

Initiation risk review

Read-only

Identifies risks in initiation materials before approval and produces structured review comments.

Initiation change assistant

Approval required

Drafts change proposals based on actual project progress and updates initiation information after approval.

R&D material issue review

Approval required

Reviews R&D material requisitions and issue details, and submits reversals for materials not used for R&D.

External AI tools

Approval required

Employees query data, handle approvals and enter time in tools such as Doubao, WorkBuddy and Qwen.

AI assistanceSuggestions within specific operations, effective after a person confirms them

Voice time entry

Human confirmed

Generates a time entry draft from a spoken account of the day’s work, which the employee confirms before submitting.

Change approval assistance

Human confirmed

Summarizes the changes and potential risks for the approver, who makes the approval decision.

Control rule drafting

Human confirmed

Turns management requirements into system rules, which are published after review.

Smart data import

Human confirmed

Matches Excel headers to system fields automatically, reducing preparation work before import.

R&D material identification

Human confirmed

Assesses the R&D relevance of material requisitions to help separate R&D use from production use.

MCP access

Use R&D Manager in Doubao and WorkBuddy

Jouyun R&D Manager provides a standard MCP service. Without logging into R&D Manager, employees complete daily work such as data queries, approvals and time entry in common AI tools such as Doubao and WorkBuddy, with the same permissions they hold in the system.

Queries and analysis

Projects, cost collection, R&D expenses as a share of revenue, company overview and term explanations.

Personal tasks

Attendance, time entry heatmap, login history, pending approvals and the task center.

People and organization

Queries and maintenance of staff, departments and positions, with row-by-row pre-checks before Excel batch imports.

Help Q&A

Answers how-to questions based on the system help center.

Authorization and controls

  • OAuth 2.1 standard authorization
  • Tiered access tokens with periodic expiry
  • Tools and data scope filtered by the user’s permissions
  • Limits on call frequency
  • Every call recorded in the run logs

Compliance checks

Every finding states its evidence type

Project compliance checks go item by item through initiation, hours, expenses, materials and policy basis, and findings are presented in four categories. The server verifies that each cited source actually exists, and findings without evidence are not treated as confirmed.

  • Evidence found

    The cited sources exist, are within the authorized scope and have passed server-side citation checks.

  • Data gap

    Information needed for the judgment is missing or insufficient, so it cannot be reliably confirmed and needs to be supplemented.

  • Data conflict

    Two or more sources contradict each other about the same fact, and a person needs to check.

  • To be verified

    Not enough valid source citations have been obtained yet, so the finding is not presented as evidence found.

AI panoramic diagnosis

5 dimensions, 200 atomic checks

Panoramic diagnosis analyzes R&D projects across five dimensions (initiation, staff, costs, results and planning), breaking 15 categories of analysis into 200 atomic checks. Rule-based results and AI analysis results are grouped separately, can be filtered by level and keyword, and link to the relevant project page or register.

  • Initiation: nature of R&D activities, innovation rationale, project similarity, fit of the R&D mode
  • Staff: fit between roles and project stages, relevance of member duties, quality of work logs
  • Costs: relevance of materials, relevance of assets, expense explanations
  • Results: process records, alignment of results with goals, support from products and IP
  • Planning: reasonableness of the schedule, budget fit
Risk diagnosis report and remediation plan

A PDF report can be generated once the diagnosis is complete. Each remediation measure cites the number of the corresponding risk. When the model is unavailable, the content is generated from rules and the report states its source.

  • Risk overview
  • Rule and AI details
  • Remediation plan
  • Remediation tracking sheet

AI initiation

Compliant from initiation

Compliance checks move forward to the initiation stage. AI works step by step from an R&D direction to a complete initiation plan, and each step can be adjusted manually.

  1. 1
    Suggest R&D directions

    Recommends possible R&D directions based on the company profile and existing projects.

  2. 2
    Refine the R&D concept

    Turns a business description into a structured R&D concept.

  3. 3
    Build the R&D profile

    Suggests the technical field, R&D mode and key resources to involve.

  4. 4
    Generate the plan in 5 sections

    Generates the core content, R&D stages, resources, budget and IP, and the main text in turn.

Plans cite only staff, equipment and materials that already exist in the system. Anything without supporting evidence is marked “To be added” for a person to complete.

Initiation collaboration

On the initiation page, users discuss the plan and revise the draft with Ouxiaojiu. Changes are first presented as candidates and are written only after the user confirms them.

Discuss the plan and revise the draft

Two modes are used for discussing the initiation approach and for editing the draft content. Basic information, stages and budget can be quoted into the chat in full or in part.

Field-level preview and undo

Suggested changes are previewed field by field, applied together once confirmed, and can be undone after they are applied. If the data has since been updated by someone else, outdated suggestions do not overwrite the newer data.

Company R&D processes

Projects are bound to the company’s preset R&D process templates. At initiation, R&D stages and document checklists are carried over, and the referenced template version is fixed.

R&D domain ontology

Shared R&D domain semantics

The R&D domain ontology defines the relationships among R&D intent, actual activities, resource inputs, expenses, evidence, rules and conclusions across 5 bases (accounting, super deduction, HNTE, IPO and as-filed), and is shared by all agents.

60
concepts
81
relation types
38
competency questions
5
accounting bases

Cognitive envelope: facts and inferences kept apart

Facts and evidenceTraceable records from authorized sources
Inferences and conclusionsJudgments drawn from facts and rules
Unknowns and conflictsMissing information and contradictory data, listed separately

Models and deployment

Options for connecting large language models

Companies choose how to connect a large language model according to their data and cost requirements.

Jouyun Cloud Gateway

Available once activated, with no model configuration required from the company.

Direct public models

Models are called directly with the company’s own model service account.

Private models

Connects models the company deploys itself. If the local model is unavailable, the system does not switch to the cloud by default.

  • Fangxin Yue Caishui
  • Doubao
  • DeepSeek
  • Qwen
  • Zhipu
  • Kimi
  • OpenAI-compatible API
  • vLLM

Compatible with mainstream LLM vendors, and can be switched by business scenario and data security requirements.

Safeguards

Controls over permissions, process and output

R&D Manager brings the execution permissions, run process and output of AI under unified control.

Three permission levels

Read-only, human confirmed and approval required, granted per capability.

Complete records

The inputs, tools called and results of every run can be viewed in the run logs.

Output validation

Sources cited in findings are validated on the server, and sensitive data in outputs is masked.

Regression evaluation

Key AI capabilities have fixed evaluation sets. After a model or prompt change, they are evaluated before release.

Frequently asked questions

Does Ouxiaojiu modify business data directly?

No. Query capabilities do not change business data. Operations that change data require approval or are submitted only after the user confirms. Every run is recorded.

How are AI findings verified?

Every finding is labeled with its evidence type. For findings marked Evidence found, the source data and policy provisions can be opened. Findings that lack evidence are marked Data gap or To be verified and are not treated as confirmed.

What does the panoramic diagnosis report contain?

The report contains a risk overview, rule and AI analysis details, a remediation plan and a remediation tracking sheet. Each remediation measure cites the number of the corresponding risk so that it can be followed up item by item.

How can companies with strict data security requirements deploy it?

Private deployment is available: R&D Manager runs on the company’s own servers and connects to a model the company deploys itself. If the local model is unavailable, the system does not switch to the cloud by default.

Which data can external AI clients access once connected?

When connected through MCP, the client acts with the employee’s own identity, and the available tools and data scope match that person’s permissions in R&D Manager.

Book a demo of the AI capabilities

A product consultant will demonstrate Ouxiaojiu’s capabilities using the company’s actual business scenarios.