Case 01 · Applied AI / Operations

TikTok Shop AI Agent Team Operations App (PANTHEON)

TikTok Shop operations.
One integrated app.

Problem
Scripts and operating records were scattered across tools.
Decision
Connected 14 AI roles through approvals, execution and reports.
Verified outcome
Integrated nine work areas into one app, with documented personal operation.
My role
Product & workflow design · AI development lead · Operator
Operating environment
Windows desktop + owner-only mobile web app
Desktop operations connecting evidence, approval, execution and reports, with mobile visibility
A diagram of the implemented scope. Actual app screens below are redacted for public viewing.
Work areas
09
Menu tabs
39
AI roles · 1 lead + 13 specialists
14
Registered job definitions
27
Actual app screens ↗

01 / Actual app screens

From system structure
to the working interface.

Actual screens from the app I operate. Operational values, dates, accounts, internal links and detailed criteria are redacted for public viewing.

02 / THE WORKSPACE

Across TikTok Shop,
one place to operate.

I connected scripts and operating records through role assignments, task controls, approvals and reports. The workspace links the information needed for a decision to the action that follows.

01TodayBriefing & decisions

A summary of metrics, inventory and role-level status to identify what needs attention.

02InboxApprovals & intervention

Separate approvals, questions, execution results and tasks requiring a person.

03PlanCampaigns & schedules

Creator invitation recommendations, stock and campaign schedules across weekly, monthly and seasonal views.

04InventoryForecasts & purchasing

Stock depletion, incoming supply, purchasing and shipment stages, plus warehouse capacity and supporting evidence.

05PerformanceSales, creators & content

Channel, product, video and creator analysis, sample funnels and official advertising data queries.

06MarketCompetitive context

Competitor snapshots, ingredient and price comparisons, and creator overlap analysis.

07ReportsResults into reporting

Weekly, monthly and campaign packs, target tracking and forecasts compared with actuals.

08TeamRoles & execution

Role-based tasks, schedules, reports, progress, responsibilities and memory.

09ConnectionsData confidence

Connection and login status, source reliability, error causes and suggested actions.

03 / FROM DECISION TO RECORD

The work continues
after the button.

Preview, approval and result tracking belong to the same workflow. Each flow distinguishes implemented code from evidence of actual operation.

01

Creator invitations & seeding

Execution records confirmed
  1. 01Stock & product input
  2. 02Target criteria
  3. 03Preview & approval
  4. 04Create or send
  5. 05Review results

Code and real runs following approval are recorded. By default invitations are only prepared; a person presses the final start. Sending stops if the verified copy file is missing, and past applicants are excluded by default. Some portal-originated registration and multi-campaign logging remain follow-up work.

02

Inventory & purchasing

Code connected · records found
  1. 01Refresh stock
  2. 02Check supply
  3. 03Preview quantities
  4. 04Update sheets

Execution paths and inventory-history tab creation are documented.

03

Content feedback

Operating record
  1. 01Select targets
  2. 02Approval
  3. 03Video review & DM batch
  4. 04Stop or resume
  5. 05Report

Off-hours night and weekend batches are started from the app’s start button. AI judges only one step each time — the video review; lookup, sending and logging are repeated by code under the same rules. Reviews run inside a subscription session, so the review step adds no per-call cost. The share of all messages sent through the app was not counted separately.

04

Content curation & sharing

Dedicated flow · posting history
  1. 01Select & review
  2. 02Bilingual draft
  3. 03Edit & approve
  4. 04Publish
  5. 05Record

VOW connects candidate selection, drafts, posting and records. Prior posting and recording history is also documented.

05

Performance management

Code connected · update history
  1. 01Refresh content & sales
  2. 02Analyze
  3. 03Report

Content, calendar and sales updates connect to scheduled and app-triggered jobs, with dashboard update history.

06

Operations management

Execution ledger confirmed
  1. 01Briefing
  2. 02Run assigned work
  3. 03Approvals & errors
  4. 04Reports & mobile

Results, errors and pending approvals persist in the report inbox and can be viewed in the owner-only mobile app.

04 / ROLE-BASED OPERATIONS

14 defined AI roles.
Managed in the same app.

One coordinator and 13 specialist roles connect organization settings, the app and operating data. Tasks are assigned to roles, with status and results flowing back into reports.

Operator · goals, priorities, approval
01Coordinator

13 specialist roles

  • 01Inventory & purchasing
  • 02Seeding
  • 03Content feedback
  • 04Metrics & settlement
  • 05Publishing
  • 06Competitor research
  • 07Development
  • 08Merchandising
  • 09Monitoring
  • 10Advertising
  • 11Partnerships
  • 12Logistics
  • 13Collection & connections
CODE

Repeatable execution

Collection, aggregation and defined procedures run through code and scheduled jobs.

AI

Interpretation & drafts

Selected judgment paths use Claude and Codex for interpretation and drafting.

HUMAN

Criteria & accountability

A person handles product selection, approval for external sends and posts, and direct login where required.

The organization groups code, scheduled jobs and selected AI judgment paths by role, with tasks and reports connected in the app.

05 / DESIGN × IMPLEMENTATION

Business judgment,
built into a product.

I defined the work, the automation boundaries, the decision owners and the review criteria, then led implementation with AI.

PHOSYND

What I led

  • Product purpose and business scope
  • TikTok Shop procedures and decision criteria
  • Role assignments, priorities and automation boundaries
  • Interface review, feedback and revision direction
  • Product, send and publishing approvals; direct operation
AI COLLABORATION

What AI helped implement

  • Technical research and design proposals
  • React/Electron interfaces and Python collectors
  • API, MCP and existing automation integration
  • Tests, reviews, bug fixes and security improvements
  • Deployment updates, documentation and user guides

The implementation layers

DESKTOP

Electron · React · TypeScript

Node.js · Vite · TanStack Query · Zustand · ECharts

DATA & EXECUTION

Python · SQLite

PowerShell · Windows Task Scheduler · Claude CLI · Codex CLI

MOBILE

React · Vinext

An owner-only web app syncing desktop status and results

Different sources, different connections
API · MCP · connectors
Creator platforms, Google Sheets/Drive/Calendar, Notion, Discord webhooks and official TikTok Ads MCP
Consoles · exports · existing pipelines
TikTok fulfillment and affiliate centers, ERP and market analysis sources
Local sources · execution records
SQLite ledger, task logs, organization settings, Obsidian and scheduled-run records

Official TikTok Ads MCP integration is distinct from full TikTok Shop Open API integration.

06 / FAILURE → RULE

Every failure
became a rule.

Problems found in operation were not left to human attention. They became rules in the app, so the next run stops or checks on its own.

01

No copy, no send

After an invitation-card copy incident, sending now stops entirely when the verified copy file is missing.

02

Past applicants excluded by default

A run re-invited creators who had already applied; excluding past applicants became the default.

03

Elements, not screen coordinates

A coordinate-based click once changed the wrong cell. Automation now targets elements directly and re-reads the record after writing to verify it.

04

Routing around a blocked integration

When direct official-API access was blocked by account conditions, I switched to official-center exports and browser paths, and put per-source freshness at the top of every screen.

05

Empty never looks like zero

Screens state why data is missing — “ledger not started”, “official source not connected” — and purchase orders without invoices are not counted as inbound stock.

07 / OPERATING CHANGE

A different way
to operate the work.

Instead of locating and running separate scripts, I can act through roles and tasks, with approvals and results in one workspace.

01

One execution workspace

Distributed scripts connected to role assignments and task controls.

02

Results that persist

Completed work, errors and pending approvals kept in reports and the execution ledger.

03

Visibility away from the desk

Desktop status and results synced to an owner-only mobile web app.

Evidence & implementation scope

Implementation scope is grounded in code, settings and operating records. Records support personal use and desktop/mobile deployment; adoption by other employees has not been quantified.

AI role scope
14 counts one coordinator and 13 specialist roles, not employees or human users. Records do not establish 14 independent LLM agents operating continuously. Assistant conversations start and end a CLI process per request.
Screens
9 areas, 39 menu tabs and 43 registered routes. Four additional routes cover questions, notes, messaging and memory. Screen registration does not establish completion of every automation.
Jobs & sources
27 job definitions, including internal, hidden and screen-specific entries; 53 stored collector task states; 60 source-status entries. These are not 60 API connections.
Code scale
77,863 lines across 514 source files in the desktop app, mobile app and collectors. Includes comments, blank lines, CSS and HTML; excludes tests, dependencies, backups, builds, data and logs.
Validation records
106 principal test files. Recorded passes of Node 203, UI 230, collectors 266 and app Python 28 come from separate checks. These are separate validation records, not a single latest-run total.
Operating boundaries
Some official-center exports, logins, final external submissions and replies require a person. Execution is not guaranteed while the PC is asleep, logged out or powered off, and automatic OS-update restarts have interrupted weekend batches.
Feature boundaries
Advertising data queries are implemented. Ad modification, automatic optimization and scheduled ad collection are outside this scope. Lark screen reading and WhatsApp notification/input handling remain in design and preparation.
Outcome interpretation
No measured before/after duration, monthly hours saved or productivity rate is claimed. Existing automation outcomes are distinguished from the effect of integrating them into the app.

Internal paths, accounts and detailed business rules remain private.