This project is an AI-powered Multi-Agent Business Decision System built for FE524 – Prompt Engineering Lab for Business Applications at Stevens Institute of Technology.
The system helps evaluate a business idea by simulating how real companies make strategic decisions.
Instead of relying on separate teams for:
- Market Research
- Technical Feasibility
- Financial Analysis
this system uses multiple AI agents to perform all three analyses automatically and generates one final business recommendation.
The final output includes:
- Go / No-Go decision
- Confidence Score
- Market analysis
- Technical feasibility report
- Financial viability report
- Consolidated risk list
- Final recommendation
The project now also includes a Streamlit web interface for live demo presentation and user-friendly interaction.
When companies plan major strategic decisions such as:
- launching a new product
- entering a new market
- building a startup
they usually need multiple teams working separately.
This creates problems like:
- slow decision-making
- inconsistent assumptions
- delayed execution
- fragmented recommendations
This project solves that by using AI agents working together inside a single LangGraph workflow.
The system is built using:
- LangGraph → multi-agent workflow orchestration
- LangChain → prompt templates and model wrappers
- OpenAI GPT-5 nano → primary LLM
- llama.cpp (fallback-ready) → local LLM backup
- PyMuPDF → PDF input parsing
- Python tools → calculator + web search simulation
- Streamlit → web-based interactive UI for demo and presentation
Responsible for:
- Market trends
- Competitor analysis
- Demand estimation
- External risks
Tool used:
- Web Search Tool
Output:
- Market opportunity summary
- Competitor analysis
- Demand assessment
- External risk flags
Responsible for:
- Technical feasibility
- Stack recommendation
- Development effort estimate
- Technical risks
Output:
- Feasibility verdict
- Recommended tech stack
- Development effort
- Technical risks
Responsible for:
- Startup cost estimation
- ROI analysis
- Break-even calculation
- Payback period
- Financial risks
Tool used:
- Calculator Tool
Output:
- Cost breakdown
- ROI projection
- Break-even analysis
- Financial risk flags
Responsible for:
- Re-analysis when confidence is low
Used when:
- confidence score is too low
- high-risk signals are detected
- business feasibility is uncertain
This improves reliability and matches real multi-agent decision systems.
Responsible for:
- Combining all agent outputs
Final output:
- Final Verdict
- Overall Confidence Score
- Research Summary
- Technology Summary
- Finance Summary
- Consolidated Risk List
- Final Recommendation
- PDF input support
- TXT input support
- GPT-5 nano integration
- OpenAI API connection
- Local LLM fallback-ready structure
- LangGraph multi-agent execution
- Web Search Tool for Research Agent
- Calculator Tool for Finance Agent
- Parallel/branching workflow
- Conditional retry routing
- Low-confidence retry node
- Shared typed state
- Final synthesis node
- Rubric-based scoring
- Consistency check across multiple runs
- Confidence validation
- Structured output evaluation
- Streamlit dashboard
- PDF/TXT upload support
- Manual business idea input
- Real-time final report generation
- Agent-wise output display
- Evaluation metrics display
- Consistency score visualization
Multi_Agent_Busisys/
│
├── app/
│ ├── main.py
│ ├── graph.py
│ ├── state.py
│ ├── llm_client.py
│ └── __init__.py
│
├── agents/
│ ├── research_agent.py
│ ├── tech_agent.py
│ ├── finance_agent.py
│ ├── retry_agent.py
│ ├── synthesis_agent.py
│ └── __init__.py
│
├── tools/
│ ├── calculator.py
│ ├── web_search.py
│ └── __init__.py
│
├── evals/
│ ├── rubric_eval.py
│ ├── consistency_check.py
│ └── __init__.py
│
├── input/
│ └── sample_business_idea.txt
│
├── output/
│
├── streamlit_app.py
│
├── requirements.txt
├── .env
├── .gitignore
└── README.md
py -m venv .venv
.\.venv\Scripts\activatepip install -r requirements.txtCreate .env
OPENAI_API_KEY=your_api_key_here
LLM_BACKEND=openai
OPENAI_MODEL=gpt-5-nanopy -m app.mainThen enter:
input/sample_business_idea.txt
or any custom PDF/TXT business idea file.
streamlit run streamlit_app.pyThis opens a browser-based dashboard where users can:
- Upload PDF/TXT files
- Type business ideas manually
- Run complete multi-agent analysis
- View final reports
- See all agent outputs
- Check rubric scores
- View consistency check results
This version is ideal for final project presentation and professor demo.
Each agent is scored based on required outputs.
Example:
- Research Agent → market, competitors, demand, risk
- Finance Agent → cost, ROI, break-even, payback
The same input is run multiple times.
The system compares:
- recommendation quality
- confidence stability
- repeated conclusions
This validates output consistency.
Example result:
Consistency Score: 87.5%
Final Verdict:
Proceed with campus-first MVP
Confidence Score:
7.5 / 10
Break-even:
55,437 orders/month
ROI:
523%
Payback Period:
18.1 months
| Course Topic | Implementation |
|---|---|
| Few-shot prompting | Agent prompt design |
| Chain-of-thought | Synthesis reasoning |
| Tool use | Calculator + Web Search |
| Open models | llama.cpp fallback |
| LangGraph | Multi-agent orchestration |
| Evaluation | Rubric + Consistency |
| Prompt Engineering | Zero-shot + Few-shot + CoT |
| Streamlit UI | Live interactive demo |
- Sri Harsha Chinta
- Kamakshi Padma Vritika Manadapaka
- Siddharth Raju Vysyaraju
- Sri Lasya Siripurapu
This project demonstrates how modern AI agents can support real-world business strategy decisions using:
- structured reasoning
- tool calling
- evaluation systems
- multi-agent orchestration
- interactive Streamlit presentation layer
rather than simple single-prompt LLM outputs.
It is designed to reflect realistic enterprise decision-making workflows and present them in a professor-level production-ready format.