Enterprise AI Copilot

Enterprises have hundreds of products new hires take weeks to learn. MyAgent eliminates that curve an AI copilot that makes anyone an expert from day one.

Comment

Inspiration

In enterprise retail, new employees face an overwhelming reality: hundreds of products, multiple service verticals (energy, logistics, digital services), and complex operational procedures they need to master before they can serve customers effectively. The typical learning curve is 2-3 weeks — during which the business loses revenue and customers lose patience.

We asked ourselves: What if an AI copilot could eliminate that learning curve entirely?

What it does

MyAgent is a multi-agent AI copilot that lets any employee — even on their first day — operate, sell, and advise like a seasoned expert. It features:

  • 7 specialized agents (Energy, Logistics, Support, Visual, Analytics, Society, Supervisor)
  • 20 MCP tools across 5 servers for real transactional execution
  • Agent Society: a multi-agent strategic debate where 5 AI agents (Sales, Marketing, Operations, Finance, Moderator) argue and converge on growth strategies using real business data
  • Persistent cross-session memory with intelligent forgetting
  • Visual analysis of bills and documents via Qwen VL
  • 7 languages with automatic detection
  • Intelligent Model Router with automatic failover across 30+ Qwen Cloud models

How we built it

  • Backend: FastAPI + LangGraph for multi-agent orchestration
  • Frontend: Next.js 14 + Tailwind CSS with SSE streaming for real-time workflow visualization
  • LLMs: Qwen Cloud (qwen3.6-flash, qwen-vl-max, text-embedding-v4) with automatic model failover
  • Protocol: MCP (Model Context Protocol) — 5 servers, 20 tools
  • Data: PostgreSQL + pgvector (RAG) + Redis (cache) + OSS (storage)
  • Observability: LangSmith + Alibaba Log Service (SLS)
  • Infrastructure: Alibaba Cloud ECS + VPC + Security Group
  • Security: Dual-layer guardrails (regex + ML content filter via Qwen)

Challenges we faced

  1. Model quota management: Free tier models exhaust quickly during development. We solved this by building an Intelligent Model Router that automatically falls back through 16 models per role — ensuring zero downtime regardless of individual quotas.

  2. MCP subprocess isolation: MCP servers running as subprocesses couldn't access environment variables. We implemented an in-process fallback transport that maintains the same MCP protocol interface.

  3. Agent Society consistency: Making 5 AI agents genuinely disagree (not just agree with each other) required carefully crafted personas with conflicting priorities and explicit instructions to challenge other agents' proposals.

  4. Multilingual in continuation mode: When the supervisor is skipped for follow-up messages, language detection had to be added at the graph level to maintain response language consistency.

Built With

Share this project:

Updates