💡 Inspiration & Vision
My inspiration for this project began with a question: How can we evolve standard e-commerce sites from static catalogs into intelligent, conversational partners? I saw an opportunity to build more than just an AI feature; I set out to create a truly dynamic, manageable, and scalable AI agentic ecosystem. My goal was to design a platform that could bring intelligent assistants to any website in a non-invasive way, all orchestrated on the power and scalability of GKE.
✨ What It Does
This project transforms the standard Online Boutique into an interactive, personalized shopping experience powered by a team of AI agents. The core of my submission is a complete, cloud-native platform that includes:
- An Injectable AI Assistant Widget: A single
<script>tag, hosted on GCS, adds a floating AI assistant to the Online Boutique product pages. It offers real-time, conversational advice, powered by a Preact and TypeScript frontend. - A Dynamic & Asynchronous Agent Backend: Running on GKE Autopilot, a
ProxyAgentintelligently handles user conversations. For complex tasks, it delegates them asynchronously via Pub/Sub to a pool of specialist agents, ensuring the user experience is always fast and responsive. - A Real-time Management Dashboard: A comprehensive Vue 3 frontend that allows for the dynamic creation of new AI agents, management of the RAG knowledge base, and live monitoring of all AI tasks and their token consumption via a real-time connection to Firestore.
🏗️ How I Built It
This project is a cloud-native application built from the ground up on Google Cloud, using a unique Human-AI collaborative development methodology.
- Backend: The core is an event-driven system of Python microservices built with FastAPI and CrewAI. This system runs on a GKE Autopilot cluster, leveraging Workload Identity for secure, keyless access to GCP services. Pub/Sub acts as the Message Coordination Platform (MCP) for asynchronous Agent-to-Agent (A2A) tasking, and Firestore provides real-time state persistence.
- Intelligence: All AI capabilities are powered by Google's Gemini models via Vertex AI. I implemented a RAG (Retrieval-Augmented Generation) pipeline using Vertex AI Vector Search to ensure all agent responses are grounded in factual, up-to-date information crawled from specified websites.
- Human-AI Collaboration: This project was a true partnership. As the human architect (Aaron Chuang), I was responsible for the core vision, architectural decisions, and final code review/refactoring. My AI partners (Gemini 2.5 Pro and Claude 3 Sonnet) acted as a pair-programming team, generating the majority of boilerplate code, writing technical specifications, and assisting in debugging complex cloud-native issues.
🚧 Challenges & Architectural Pivots
The primary challenge was architecting and implementing a highly elastic, manageable AI agent platform within the hackathon's tight timeframe.
My initial concept was a fully asynchronous model where a central Orchestrator handled all requests. While scalable, I quickly realized this could introduce unacceptable latency for real-time user conversations.
Therefore, I made a critical architectural pivot to a more sophisticated Proxy/Orchestrator pattern. A lightweight Proxy Agent now handles all user-facing conversations for instant interaction, while complex tasks are asynchronously delegated. This refactoring cost precious time and meant I couldn't fully implement the multi-agent delegation feature, but it was the right decision to prioritize a seamless user experience and build a more robust foundation. This hands-on process of building, testing, and making tough architectural decisions under pressure was the most significant challenge and learning experience.
✅ Accomplishments I'm Proud Of
Frankly, what I'm most proud of is successfully building and deploying this complete, end-to-end cloud-native AI platform from scratch. I didn't just build a single feature; I built an entire ecosystem. This includes:
- A real-time monitoring dashboard built with Vue 3.
- A non-invasive, injectable frontend widget built with Preact.
- A scalable, asynchronous AI backend orchestrated on GKE Autopilot.
I am particularly proud of successfully navigating and resolving complex, real-world cloud engineering challenges, especially in the GKE Ingress networking layer. Furthermore, validating the Human-AI collaborative development paradigm was a game-changer, proving that such a sophisticated system could be architected and largely built by a solo developer in a matter of days.
🚀 What's Next
The current platform provides a robust foundation for a truly autonomous and scalable AI workforce. My roadmap is focused on bringing the "AI as a Company" concept to life:
- Full Implementation of the "AI HR Department" (Dynamic Agent Management): Completing the automated pipeline where the
OrchestratorAgentcan "hire" new AI specialists by triggering a serverless CI/CD process to provision GKE deployments and Pub/Sub topics automatically. - Building the "Corporate IT Department" (Centralized Tool Management): Creating a UI for registering new capabilities (e.g., new APIs) as shareable "tools" that can be dynamically assigned to agents, allowing the ecosystem to learn new skills without redeployments.
- Introducing the "Strategic Planning Department" (AI-Powered Auditing): Creating a high-level "Auditor Agent" that analyzes operational data to provide proactive strategic advice, such as recommending the creation of new agents based on market trends identified in the RAG knowledge base.
- Enhancing Long-Term Conversational Memory: Evolving beyond single-session memory by leveraging a persistent store like Firestore or Cloud Memorystore to provide a truly continuous and personalized user experience across multiple visits.
Log in or sign up for Devpost to join the conversation.