Skip to content
Yanyu Zhou
Portfolio
Case study 04Project management · Knowledge management · Reporting

Project Intelligence & Operating Systems

Turned fragmented project communication into continuously refreshed operational knowledge that teams and agents can reuse.

Delivered · in use

Business value

Every approved project starts with a ready workspace, and the team’s agent works from current project information instead of a stale snapshot.

My role
Process design · Automation build · Pipeline design · Agent setup · Rollout
Skills
  • Workflow automation
  • Data pipeline design
  • Knowledge management
  • Agent design
  • Project operations
Built with
  • Collaboration workspace
  • Native agent platform
  • Approved communication archive
  • Webhooks and scheduled jobs
  • AI knowledge layer

Confidentiality note. Company-specific data, configurations, workflows and diagrams in this case study have been anonymized and reconstructed for public presentation.

01

The request

From · Project teams

Project knowledge was spread across email, files and individuals. The real AI problem was not whether an agent could answer project questions — it was how the agent stays current when project information changes every day.

  • Pain point 01

    Every new project was set up by hand.

  • Pain point 02

    Project knowledge was spread across email, files and individual people.

  • Pain point 03

    An agent only knew what someone last pasted in.

02

What I built

  1. 01A project operating model: once a project is approved, a standard workspace is created automatically — project group chat, task list and project document, linked to each other
  2. 02A team agent that works inside that workspace
  3. 03A recurring knowledge pipeline that feeds approved project communication into the agent’s knowledge layer
03

Try it — approve a project, then run a refresh

Approve a sample project to see its workspace created, then run the daily refresh and watch new communication reach the agent.

How to use
  • The project and messages are fictional.
  • Notice what happens to the message that can’t be matched with certainty.
Project workspace
Spring Launch

Request submitted · waiting for approval.

Daily refresh · approved project communication

Runs once the workspace exists.

Simplified demo · fictional data · not the real system
04

How I thought about it

  1. 01

    Approved archive, not individual inboxes

    Only approved, team-shared project communication enters the pipeline. Personal mailboxes are never connected.

  2. 02

    Never guess a project match

    A thread is linked to a project only when the match is certain; anything else is flagged for a person.

  3. 03

    Deliver knowledge where work happens

    The workspace and the agent live in the team’s collaboration space, not in another tool people have to remember.

05

How it was delivered

Run as a project: from the team’s request to something people use every day.

  1. 01
    Intake

    Mapped how projects were set up and where project information actually lived.

  2. 02
    Design

    Defined the standard workspace per project type and which communication is approved to feed the agent.

  3. 03
    Build

    Built the approval-to-workspace automation, the daily retrieval pipeline and the agent’s knowledge layer.

  4. 04
    Test

    Verified matching, redaction and checkpoint behaviour before the daily schedule went live.

  5. 05
    Roll out

    Brought the team agent into project workspaces where the team already works.

06

Result

BeforeEach new project was set up by hand
AfterAn approval creates the project workspace automatically
BeforeProject knowledge spread across email, files and people
AfterA knowledge layer refreshed daily from approved communication
BeforeResearch and drafting started from zero
AfterAI-assisted timelines, research and documents for people to review
An approval now turns directly into a working project space, and teams and their agent work from current project information.
07Technical detailsFor the deeper read: From approval to workspace · How does the agent stay current? · How does the agent know? · How are models and tools used? · Automated Weekly Business Review · What remained controlled / human

From approval to workspace

  1. 01
    Approve

    A project request is approved.

  2. 02
    Create

    A workspace matching the project type is created automatically: project group chat, task list and project document, embedded in each other.

  3. 03
    Join

    The team agent joins the workspace and works where the team already works.

  4. 04
    Maintain

    The agent updates and flags the project document from new communication, task progress and chat.

  5. 05
    Close

    Each project document ends with a scorecard section, connecting project work to the measurement layer.

The scorecard at the end of each project document feeds the client and vendor intelligence layer. Case 02 →

How does the agent stay current?

Approved project communications are retrieved on a recurring basis, processed into usable context, and added to the team agent’s knowledge layer — so the agent operates from current project information rather than a static snapshot.
  1. 01
    Retrieve

    Only approved, team-shared project communication is retrieved from an archive. Individual inboxes are not connected.

  2. 02
    Process

    Messages are normalized into threads. Credentials and payment details are redacted. Threads are linked to a project only when the match is certain — otherwise marked unmatched, never guessed.

  3. 03
    Publish

    Processed threads are published as structured records and documents in the collaboration platform.

  4. 04
    Ingest

    New material is added to the team agent’s knowledge layer.

  5. 05
    Advance

    The checkpoint moves forward only after writes are verified; overlapping windows catch late-arriving messages.

How does the agent know?

  1. 01
    Persistent knowledge
    Project brief · deliverables · key decisions · core documents
    Long-term project memory, kept in the project workspace
  2. 02
    Fresh context
    Latest communication · timeline changes · approvals · progress updates
    Added through the daily refresh
  3. 03
    Task context
    Only what the current request needs
    Passed into each action — nothing broader
Knowledge grounding and recurring context ingestion — not model training.

Action layer

  • Timeline generation
  • Partner research / matching
  • Location research
  • Talent research
  • Information retrieval
  • Document support

How are models and tools used?

Native agent platforminside the project workspace
Retrieval, project-document updates and answering the team
Rules-based pipelinescheduled daily
Retrieval, normalization, redaction and publishing — deterministic, so no AI
Agent actions
Research and drafting — timelines, research, scouting — reviewed by a person
Human
Decisions, approvals and anything client-facing

Automated Weekly Business Review

A recurring company-wide operating review

A weekly workflow that distributes follow-up tasks, consolidates updates into a consistent structure, and produces a company-level operating snapshot with outstanding actions surfaced.

The weekly cycle
  1. 01
    Distribute weekly follow-up tasks
  2. 02
    Collect updates
  3. 03
    Consolidate & structure
  4. 04
    Company operating snapshot
  5. 05
    Surface outstanding actions
Outstanding actions carry into the next week’s cycle.

What remained controlled / human

  • Which sources are approved to feed the knowledge layer.
  • Ambiguous project matches — flagged for a person, never guessed.
  • Decisions and client-facing output; agent research and drafts are inputs for review.