Inspiration
Small businesses have access to a lot of data—sales, inventory, customers, expenses, and marketing—but turning that data into the right business decisions is difficult.
We wanted to build something more useful than a traditional dashboard. Instead of making a business owner analyze dozens of numbers themselves, we wanted an AI system that could act like a virtual management team.
This inspired us to create SmallBiz AI Manager: a multi-agent AI system that understands business data, researches external market trends, and turns them into practical business insights.
Our goal was simple:
Don't just show business data. Help the business understand what to do with it.
What it does
SmallBiz AI Manager is an AI-powered business intelligence and management platform for small businesses.
It combines internal business data with AI-powered analysis to provide insights across multiple areas:
Sales Agent — analyzes sales performance, products, customers, and inventory. Finance Agent — analyzes revenue, expenses, orders, and financial performance. Marketing Agent — analyzes website traffic, customer behavior, acquisition channels, and marketing performance. Trend Search Agent — searches the internet for current and upcoming trends, events, product launches, festivals, gaming releases, and other factors that could influence demand. Root Agent — acts as the central business manager, deciding which specialist agents should be involved. Display Agent — converts AI results into structured output that can be presented directly in the React interface.
For example, if an upcoming event could increase demand for a particular product category, the Trend Agent can identify the trend while the business agents provide information about current sales and inventory. The system can then combine these signals into a more informed recommendation.
The Marketing Agent can also execute actions such as sending emails when explicitly instructed by the user.
The overall flow is:
$$ \text{Business Data} + \text{External Trends} + \text{AI Agents} \rightarrow \text{Business Insights} \rightarrow \text{Action} $$
How we built it
We built SmallBiz AI Manager using React + Vite for the frontend, Flask for the backend, MongoDB for business data, and Google Gemini + Google Agent Development Kit (ADK) for the AI system.
The application contains business modules for:
Dashboard Sales Orders Inventory Expenses Reports Marketing AI Manager
The backend exposes APIs that connect the application to MongoDB and provide business information to the AI agents through tools.
Multi-Agent Architecture
Instead of using a single AI agent for everything, we created specialized agents with clearly defined responsibilities.
┌───────────────────────┐
│ USER / OWNER │
└───────────┬───────────┘
│
▼
┌─────────────────────────────┐
│ REACT FRONTEND │
│ Vite + React │
│ │
│ Dashboard │
│ Sales │
│ Orders │
│ Inventory │
│ Expenses │
│ Reports │
│ Marketing │
│ AI Manager │
└─────────────┬───────────────┘
│
┌────────────────┴────────────────┐
│ │
▼ ▼
┌────────────────────────┐ ┌────────────────────────┐
│ GOOGLE CLOUD │ │ GOOGLE CLOUD │
│ BUSINESS API ENDPOINTS │ │ AI AGENT ENDPOINT │
│ │ │ │
│ Sales API │ │ Python AI Application │
│ Orders API │ │ Google ADK │
│ Inventory API │ │ Gemini 3.5 Flash │
│ Customer API │ │ │
│ Finance API │ └───────────┬────────────┘
│ Marketing API │ │
│ Reports API │ │
└───────────┬────────────┘ │
│ │
▼ ▼
┌───────────────────┐ ┌───────────────────┐
│ DATABASE │ │ AI AGENT SYSTEM │
│ │ │ LEVEL 2 │
│ MongoDB │ └───────────────────┘
│ │
│ Products │
│ Orders │
│ Order Items │
│ Inventory │
│ Customers │
│ Expenses │
└───────────────────┘
┌───────────────────────┐
│ REACT AI MANAGER │
│ │
│ User Query │
└───────────┬───────────┘
│
▼
┌───────────────────────┐
│ AI ENDPOINT │
│ Google Cloud │
└───────────┬───────────┘
│
▼
╔══════════════════════════════╗
║ ROOT AGENT ║
║ Business Manager ║
║ ║
║ Understands user request ║
║ Routes tasks ║
║ Coordinates specialists ║
╚══════════════╤═══════════════╝
│
┌──────────────────┼───────────────────┐
│ │ │
▼ ▼ ▼
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ SALES AGENT │ │ FINANCE AGENT│ │ MARKETING │
│ │ │ │ │ AGENT │
│ Sales │ │ Revenue │ │ Analytics │
│ Products │ │ Expenses │ │ Customers │
│ Customers │ │ Orders │ │ Campaigns │
│ Inventory │ │ Financial │ │ Email │
└──────┬───────┘ │ Analysis │ │ Execution │
│ └──────┬───────┘ └──────┬───────┘
│ │ │
└──────────────────┼───────────────────┘
│
▼
┌─────────────────────┐
│ TREND SEARCH │
│ AGENT │
│ │
│ Current Trends │
│ Product Launches │
│ Festivals │
│ Gaming │
│ Sports │
│ Consumer Trends │
└──────────┬──────────┘
│
▼
┌──────────────────┐
│ GOOGLE SEARCH │
│ External Web │
└────────┬─────────┘
│
▼
┌─────────────────────┐
│ BUSINESS │
│ RECOMMENDATION │
│ │
│ Sales + Finance │
│ Marketing + Trends │
│ Inventory + Data │
└──────────┬──────────┘
│
▼
╔══════════════════════╗
║ DISPLAY AGENT ║
║ ║
║ Structured Output ║
║ Tables ║
║ Charts ║
║ Cards ║
║ UI-ready Response ║
╚══════════╤══════════╝
│
▼
┌─────────────────────┐
│ REACT FRONTEND │
│ │
│ AI Response │
│ Charts / Tables │
│ Recommendations │
└─────────────────────┘
Each specialist agent uses tools to access the information it needs rather than inventing business data.
For example, the Sales Agent can retrieve sales and inventory information, while the Marketing Agent can retrieve website analytics and execute email actions when requested.
The Trend Search Agent uses Google Search to gather current information from the internet, allowing the system to consider real-world events alongside internal business data.
Finally, the Display Agent ensures that the final response is converted into a structured format that the frontend can understand and display as text, tables, charts, or other UI components.
One of the biggest challenges was making these agents work together efficiently while maintaining reliable, structured output. We designed each agent with a specific responsibility to reduce unnecessary reasoning, prevent hallucinated business information, and make the overall system easier to scale.
Challenges we ran into
Building a multi-agent AI system came with several challenges.
The biggest challenge was agent orchestration and latency. When a user asks a complex question, the Root Agent may need to delegate the task to multiple specialist agents. Each handoff adds processing time, so we had to carefully design responsibilities and reduce unnecessary agent calls.
Another challenge was making AI responses reliable. We did not want the system to invent sales numbers, customer information, financial metrics, or market trends. We therefore designed our agents to rely on actual tools and business data and to clearly distinguish facts from assumptions and predictions.
Integrating Google Search with Gemini and ADK was another technical challenge, particularly because Google Search is a built-in server-side tool and requires the correct tool configuration.
We also had to solve the challenge of connecting unpredictable AI output to a structured React interface. This led us to introduce a dedicated Display Agent, which converts the final AI response into a predictable structure that our frontend can render.
Finally, connecting AI agents to real actions—such as sending marketing emails—required careful separation between recommendations and execution so that the system does not perform actions without the user's request.
Accomplishments that we're proud of
We are proud that we built more than just an AI chatbot.
We created a complete AI-powered business management platform with a multi-agent architecture where each agent has a specific role.
Some of the accomplishments we are most proud of are:
- Built a working business management dashboard with React and Flask.
- Integrated MongoDB for real business data.
- Created specialized Sales, Finance, Marketing, and Trend Search agents.
- Built a Root Agent capable of coordinating specialist agents.
- Integrated Google Search for real-time market trend research.
- Created a dedicated Display Agent for structured frontend output.
- Connected AI agents to actual business tools and data.
- Added email execution capabilities to the Marketing Agent.
- Designed the system to distinguish facts, observations, hypotheses, and recommendations.
- Focused on preventing AI hallucinations instead of simply generating convincing-looking answers.
Most importantly, we created an architecture that can be expanded with additional specialist agents without redesigning the entire application.
What we learned
This project taught us that building a useful AI system requires much more than choosing a powerful model.
We learned how to design multi-agent systems, where each agent has a narrow responsibility and a larger agent coordinates them.
We also learned the importance of tool calling. Instead of asking an AI to guess business information, we can give it tools that retrieve actual data and allow it to interact with external systems.
Another major lesson was structured AI output. A chatbot can generate free-form text, but an application needs predictable data. Designing the Display Agent helped us bridge the gap between AI reasoning and traditional software interfaces.
We also learned that AI reliability is a system-design problem. Good prompts alone are not enough. Agents need clear responsibilities, access to trustworthy data, validation, and rules for handling uncertainty.
Finally, we learned that performance matters. A powerful multi-agent system is not useful if every simple question triggers unnecessary agents. Designing efficient delegation and minimizing handoffs became an important part of our architecture.
What's next for SmallBiz AI
SmallBiz AI Manager is only the beginning.
Our next goal is to make the system more proactive instead of waiting for the business owner to ask questions.
We want the AI Manager to monitor business activity and identify important events automatically, such as:
- Unusual changes in sales
- Products approaching low stock
- Sudden changes in customer behavior
- Marketing performance drops
- Emerging market trends
- Upcoming events that could affect demand
- Products gaining or losing popularity
We also want to improve the Recommendation Agent so it can combine sales history, inventory, customer behavior, financial information, and external trends into more sophisticated business recommendations.
Future versions could also include:
- Predictive sales forecasting
- AI-powered inventory optimization
- Automated customer segmentation
- Smarter marketing campaign generation
- More real-time market intelligence
- Proactive AI business alerts
- Demand prediction for individual products
- More natural conversational business analytics
- Stronger permission and approval systems for AI actions
Our long-term vision is to turn SmallBiz AI Manager into a virtual AI management team for small businesses—one that doesn't just tell owners what happened, but helps them understand what is happening, what might happen next, and what they can do about it.
What's next for SmallBiz AI
Our next goal is to make SmallBiz AI more proactive, predictive, and autonomous.
Instead of waiting for business owners to ask questions, we want the AI Manager to continuously monitor business activity and alert them when something important happens—such as unusual sales changes, low inventory, changing customer behavior, or emerging market trends.
Log in or sign up for Devpost to join the conversation.