BizBrain

A "Virtual CEO" or business second brain. The agent helps business owners interact with HR policies, manuals, financials, sales records, etc. to provide actionable executive insights.

Comment

Inspiration

Business leaders and founders spend hours context-switching between unstructured documentation (HR manuals, brand guidelines, meeting notes) and quantitative data spreadsheets or database ledgers (sales records, P&L statements, expense tracking). Standard LLM chatbots fail at quantitative reasoning over raw database records via simple vector search, while traditional BI dashboards lack qualitative contextual awareness.

We set out to build Virtual CEO (BizBrain): an intelligent, plug-and-play business operating copilot that unifies qualitative semantic memory with precise, structured SQL database execution. Powered by Qwen AI, Virtual CEO acts as a 24/7 business advisor capable of answering operational questions, querying financial ledgers via Text-to-SQL, and guiding strategic decisions in real time.

What It Does

Virtual CEO is built around a Triad Memory Architecture:

  1. Identity & Core Rules Layer: Injected into Qwen's system prompt to ground the agent in company OKRs, brand tone, mission, and operational boundaries.
  2. Unstructured Knowledge Vault (Vector RAG): Leverages Qwen text embeddings to index PDFs, manuals, and qualitative strategic logs for semantic search.
  3. Structured Financial & Operations Memory (Text-to-SQL): Connects Qwen to a relational MySQL database containing real-time revenue, expense, and transaction records, allowing the agent to dynamically generate and execute precise SQL queries.

Key Features:

  • Hybrid Query Routing: Qwen automatically decides whether a user prompt requires SQL database execution, vector search over qualitative documents, or direct core memory retrieval.
  • Transparent Tool Execution: Shows live agent "thoughts" and tool calls (e.g., query_financial_db, search_unstructured_vault), giving users complete visibility into how answers are derived.
  • Document & Ledger Ingestion: Built-in upload interface for business owners to ingest PDFs into the vector vault and CSVs into the relational SQL store.

How We Built It

  • AI Engine & Reasoning: Powered by Qwen-Max and Qwen-Plus via the Alibaba Cloud DashScope API, utilizing native tool-calling capabilities.
  • Embeddings: Qwen Text Embeddings for semantic document chunking and vector retrieval.
  • Backend: Node.js & Express orchestration engine providing API endpoints and agent tool handlers.
  • Frontend: Modern React (Vite + Tailwind CSS) user interface featuring real-time stream responses and interactive tool execution cards.
  • Databases: MySQL for structured financial ledgers and ChromaDB / Pinecone for vector document embeddings.
  • DevOps & Infrastructure: Fully containerized with Docker Compose for instant one-command deployment, automated via GitHub Actions for VPS hosting.

Challenges We Ran Into

  1. Multi-Store Context Routing: Teaching the LLM when to perform a relational SQL query versus a vector similarity search required refining tool descriptions and few-shot examples so Qwen reliably selected the correct data source.
  2. Safe & Accurate Text-to-SQL: Ensuring Qwen generates valid SQL queries without hallucinating table schemas required dynamically injecting schema context and strict query sanitization rules into the tool interface.
  3. Low-Latency Agent Loops: Balancing multi-step tool execution with responsive frontend streaming required fine-tuning tool payloads and state management.

Accomplishments That We're Proud Of

  • Flawless Tool Orchestration: Achieving seamless multi-step routing where Qwen can evaluate a query like "What was our net profit last month and how does our expense policy handle refunds?" by querying both MySQL and the Vector DB in a single execution flow.
  • Production-Ready DX: Delivering a complete, containerized stack that spins up with a single docker-compose up command.
  • Visual Transparency: Designing a clean UI that explicitly renders the agent's internal reasoning steps and tool execution outputs for user trust.

What We Learned

  • How to maximize Qwen AI's function-calling capability for dual-database orchestration.
  • The fundamental importance of separating quantitative (relational SQL) and qualitative (vector RAG) storage layers in business AI applications.
  • Strategies for designing modular agent skills that can be safely extended as business requirements grow.

What's Next for Virtual CEO (BizBrain)

  • Executive Agent Board: Expanding into a multi-agent system (Virtual CFO, Virtual CMO, Virtual COO) that collaborate on complex tasks.
  • Native Integrations: Direct API connectors for QuickBooks, Xero, Stripe, Shopify, and Google Workspace.
  • Proactive Insights: Automated scheduled jobs that audit business metrics weekly and alert the owner to revenue dips or unusual expenses.

Built With

Share this project:

Updates

Submission history