← Back to selected work

TrafficDeck

Fully applied AI: a browser traffic product built through a zero-code approach with a LangGraph agent team.

Context
Independent product and engineering project
My contribution
Product direction, engineering judgment, and human acceptance
Technologies
Python · LangGraph · Go · React · TypeScript · MCP · Playwright

The idea

TrafficDeck lets people redirect, modify, delay, block, and mock browser requests. The product connects a browser extension, a React dashboard, and a Go API through shared rules and contracts.

I approached its delivery as fully applied AI: use engineering experience to direct an agent team from the initial idea through implementation, testing, review, and release preparation.

The zero-code approach

My role is to define the problem, set direction, challenge decisions, and accept the result. AI agents perform the implementation work. From my side, this is a zero-code approach to product delivery, without a separate human developer role.

LangGraph orchestrates the team and its handoffs. The product itself contains real code, tests, and infrastructure; the zero-code approach describes how I lead its creation.

Meet the agent team

The workflow gives specialist responsibilities to a coordinated team: Project Manager, Architect and Architecture Critic, Tech Lead, Extension, Backend and Frontend Developers, QA Engineers, Code Reviewer, Security Reviewer, Workflow Coordinator, and Release Engineer. Infrastructure and CI roles support the delivery environment.

Product + architectureAgent developersQA + code + security reviewHuman acceptance

Durable state, scoped responsibilities, and review loops make the work traceable and resumable. Human acceptance remains the final decision.

The decisions that matter

  • Keep the contracts explicit. The extension, dashboard, and API need a shared interpretation of rules and their behavior.
  • Make agent work resumable. Durable execution state makes it possible to understand what finished, what failed, and what needs attention.
  • Bound the rework. Review loops need an exit condition and a clear owner when an automated process cannot resolve a problem.
  • Keep people in the process. Human review and acceptance remain part of the workflow.

System design for an agent team

Agent orchestration brings system design questions into the delivery process: where state lives, which role owns a task, how a failed step resumes, and when a review loop must stop. I make those boundaries explicit in the LangGraph workflow.

What this project shows

Applied AI can be an operating model for building a product, spanning planning, development, quality, and delivery. TrafficDeck is my independent, ongoing exploration of that model, guided by 15 years of engineering experience.