MSPixelPulse — Building my own AI engineering system

How Mahak Patel built a private AI engineering workspace with specialized agents, reusable skills, owner access, scoped tools and human review for MSPixelPulseAgency.

My role: product design, full-stack development and AI systems engineering.

From repeated chats to reusable context

A private workspace to preserve project context, specialized responsibilities, repeatable engineering workflows and human review.

Agents

20 registered agents, audited October 3, 2026. Defined responsibilities do not imply autonomous execution.

Build

  • Web engineer
  • Application engineer
  • UI/UX designer
  • Architecture engineer
  • Data & integrations

Quality

  • QA engineer
  • Accessibility engineer
  • Security & privacy
  • Performance & SEO
  • Code reviewer

Ship

  • Release engineer

Research & content

  • Research agent
  • Content strategist

Growth

  • SEO strategist
  • Analytics & conversion
  • Lead research
  • Outreach & CRM
  • Inquiry triage

Intelligence

  • AI intelligence & capability scout

Skills

45 registered workflows; availability depends on model, connection and permission requirements.

  • repository-audit: Establish the project, architecture and evidence before changing code.
  • ui-ux-design: Translate requirements into a coherent interface and interaction direction.
  • responsive-qa: Check real layouts across desktop, tablet and narrow mobile screens.
  • security-privacy-review: Review trust boundaries, sensitive data and access decisions.
  • test-and-verify: Validate the result and report what was actually tested.
  • safe-production-release: Prepare release evidence and retain explicit human approval.

How a task flows

  • Define the goal: I set the outcome, project and scope.
  • Route the request: The routing table suggests a specialist and relevant skills.
  • Load context: Selected knowledge and skill instructions supply the working context.
  • Work with tools: The server checks capabilities, connections and action policy.
  • Validate evidence: Checks and task events make the outcome reviewable.
  • Review and approve: I decide what to accept and approve external actions.

Architecture and owner access

React and Express separate the owner interface from the private server runtime, policy checks and hosted PostgreSQL persistence. Google verifies the owner during enrollment or recovery; routine access uses a passkey.

Technology

Product

  • React
  • TypeScript
  • Vite
  • Node.js / Express
  • PostgreSQL
  • Better Auth
  • WebAuthn
  • Vercel

Agent system

  • OpenAI Agents SDK
  • Markdown charters
  • Agent Skills
  • Routing registry
  • Shared knowledge
  • Tool policy

Connection implementations

  • GitHub
  • Vercel API
  • Gmail
  • Search Console
  • Google Analytics 4
  • Google Ads

Human control

Goals, product decisions, sensitive actions and production releases remain subject to human judgment. Paid model execution is separately gated. Provider implementations do not imply authorized live connections.

What I learned

  • Context before headcount: Adding another agent is easy. Giving it useful context, a clear responsibility and a way to verify its work is the real design problem.
  • Orchestration is product design: Routing, permissions and clear status messages shape the experience as much as the model does. A saved task should never look like completed work.
  • Correctness still needs judgment: Automated checks can show evidence. I still need to decide whether the result solves the actual problem and whether it should reach production.

Explore the separate CanSTEM Enrollment Portal project

Back to Work