Enterprise AI Agent Platform
A multi-tenant platform that combines custom agents, tools, and knowledge bases to build and operate AI assistants specialized for each company's operations.
Challenge
A multi-model AI agent platform was needed to leverage corporate knowledge and automate business operations. The design had to meet security, multi-tenancy, and scalability requirements while enabling no-code agent building.
Solution
Built a clean architecture-based backend and secure infrastructure on GCP. Integrated PydanticAI for agent execution, LlamaIndex + pgvector for RAG pipeline, and RestrictedPython for code sandboxing.
Result
Built a full-stack platform with 150+ frontend, 80+ backend, and 50+ infrastructure files.
Team
2 members, 4 months
Design, implementation, infrastructure, operations
Role
Responsible for everything from design to implementation, infrastructure, and operations.
Led frontend (Next.js), backend (FastAPI), and infrastructure (Terraform/GCP) end-to-end.
Tech Stack
Key Features
Real-time AI chat via SSE: 6 types of streaming events displaying thought processes and tool execution status live
Multi-model support: Switch between OpenAI / Anthropic / Gemini through a unified interface
No-code agent builder: Freely configure system prompts, tools, and models
RAG knowledge base: Vectorize documents and improve AI response accuracy with semantic search
Code sandbox with RestrictedPython: Secure user code execution environment
Multi-tenancy: Structurally prevent cross-tenant access with workspace_id filtering across all tables
Technical Highlights
Clean Architecture
Unified dependency direction from domain → application → infrastructure → api, enabling LLM provider switching with only infrastructure layer changes.
Enterprise Security
Restricted backend to internal traffic only, implemented dual authentication with OIDC + user tokens, DB isolation via VPC, and secret management via Secret Manager.
Cost Optimization
Reduced development environment costs to 2-5% of production through Cloud Run scale-to-zero and environment-specific resource isolation. All infrastructure managed as code with Terraform.