Keith Chiu

← Lab · Case study · AI operating model

My AI Operating Stack

The architecture behind how I work: AI agents, connected data, scheduled automations and a push-to-live pipeline, with a human approval gate wherever it matters.

Architecture

AI operating stack architecture Five lanes: Operator (Keith, optional Claude Code), AI agents (Claude, scheduled agents, browser agent, cloud workspace), Data and context (Google Drive, Gmail, Calendar and Sheets, persistent memory), Pipeline (GitHub, Cloudflare Pages CI/CD) and Delivery (Cloudflare DNS, keithchiu.net, global CDN, email). Changes pushed from the workspace to GitHub are automatically deployed by Cloudflare Pages to the CDN serving keithchiu.net. Operator AI agents Data & context Pipeline Delivery KeithBrief · review · approve Claude CodeOptional local dev ClaudePlan · write · test Scheduled agentsRecurring research Browser agentActs with approval Cloud workspaceShell · Python · Git Google DriveDocs · CVs · notes Gmail · Calendar+ Sheets connectors Persistent memoryContext · preferences GitHubPrivate repo · history Cloudflare PagesAuto-deploy on push Cloudflare DNSDomain routing keithchiu.netCV + Lab Global CDNHTTPS · rollback EmailGoogle Workspace

AI agent Persistent data Ephemeral or optional Platform / public

What runs on its own

Trigger · git push

Push-to-live deployment

Every approved change is committed to version control and deployed worldwide in under a minute. Every release is kept, so any version can be restored in one click.

Trigger · schedule

Scheduled agents

Recurring research runs on a timetable without me starting it, and the results are waiting when I open my laptop.

Trigger · request

Connected workspace

Agents read and write across Drive, Gmail, Calendar and Sheets, so work starts from real context instead of copy-paste.

Trigger · request + approval

Browser agent

Routine web admin (dashboards, settings, DNS) is done by an agent in the browser, pausing for my sign-in and approval at each sensitive step.

Proof it works

Brief to live
1 day
site + 4 tools
Push to live
< 1 min
fully automated
Maths tested
4 / 4
tools, before release
Tool inputs collected
0
in every Lab tool

Guardrails

Accuracy

  • Verify, don't trust. Statistical outputs are cross-checked against an independent reference library before release.
  • No invented facts. Gaps are flagged for a human to confirm, or cut.
  • Edge cases first. Tools are tested with bad inputs and boundary cases, not just the happy path.

Control & privacy

  • Humans hold the keys. Sign-ins, approvals and irreversible changes stay with me.
  • No employer data. Lab defaults are illustrative; nothing confidential is used.
  • Least exposure. Private repo, browser-only tools, and pages kept out of search indexes.

The same model at work

In my e-commerce role I run the same pattern across AI-assisted analytics and reporting, AI translation for new-language rollouts, AI customer service and automated executive reporting: a clear brief, agents for speed, a human for judgement, and a verification gate before anything reaches a customer or an executive.

Built by Keith Chiu with AI-assisted development (Claude). See the output: A/B Test Calculator · Funnel Uplift Modeller · KPI Anomaly Radar · Growth Strategy Builder.