AI Multi-Agent Business Decision System

Project Overview

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 evaluates business ideas by simulating how real companies make strategic decisions. Instead of relying on separate teams for market research, technical feasibility, financial analysis, and compliance review, this project uses multiple AI agents inside a single LangGraph workflow to generate one final business recommendation.

The final output includes:

Go / No-Go decision
Overall confidence score
Market analysis
Technical feasibility report
Financial viability report
Regulatory compliance context
Consolidated risk list
Final recommendation
Downloadable Word report from the Streamlit interface

Business Problem

When companies evaluate a new product, startup idea, or market-entry strategy, they usually need multiple teams working separately:

Market research team
Technology team
Finance team
Compliance or regulatory team
Strategy team

This creates common business problems:

Slow decision-making
Inconsistent assumptions across teams
Fragmented recommendations
Delayed execution
Difficulty comparing feasibility across market, technology, and finance

This project solves the problem by using AI agents that work together in a coordinated workflow. Each agent specializes in one decision area, and the synthesis layer combines the outputs into a single final report.

System Architecture

The system is built using:

LangGraph for multi-agent workflow orchestration
LangChain for model wrappers and embedding integrations
OpenAI GPT-5 nano as the primary LLM
OpenAI Embeddings for vector search
ChromaDB for local vector storage
PyMuPDF for PDF parsing
Python tools for calculator and market context utilities
Streamlit for the interactive web interface
python-docx for downloadable Word reports
llama.cpp fallback-ready structure for optional local LLM support

High-Level Workflow

Business Idea Input
        ↓
PDF/TXT Parsing or Manual Input
        ↓
Domain Detection
        ↓
Research Agent
(Market RAG + Market Context Tool when domain matches)
        ↓
Technology Agent
(System Architecture + Implementation Plan)
        ↓
Finance Agent
(Calculator Tool + Unit Economics)
        ↓
Regulatory RAG
(State-Specific Compliance Retrieval)
        ↓
Synthesis Agent
(Final Business Decision Report)
        ↓
Evaluation Layer
(Rubric + Consistency Check)
        ↓
Downloadable Word Report

Agent Workflow

Agent 1 — Research Agent

The Research Agent evaluates the business idea from a market and customer perspective.

Responsibilities:

Market opportunity analysis
Target customer identification
Customer pain-point analysis
Market trend analysis
Competitor analysis
Demand assessment
Pricing and adoption insights
External risk identification
Go-to-market recommendation
Confidence score generation

Tools and retrieval:

Market RAG for domain-specific grocery/startup context
Market context tool
Domain-aware fallback for non-grocery ideas

Output:

Structured JSON research analysis
Market opportunity summary
Competitor analysis
Demand assessment
Risk flags
Confidence score

Agent 2 — Technology Agent

The Technology Agent evaluates whether the business idea is technically feasible.

Responsibilities:

Technical feasibility assessment
Recommended technology stack
System architecture design
Frontend and backend requirements
Database and cloud infrastructure recommendations
AI/ML architecture planning
Deployment and monitoring plan
Technical risk analysis
Development timeline and MVP planning

Output includes:

Feasibility verdict
Project requirements
Systems architecture
Development effort
Operational requirements
Technical risks
Confidence score

Agent 3 — Finance Agent

The Finance Agent evaluates financial feasibility using extracted parameters and calculator-based analysis.

Responsibilities:

Startup cost estimation
Revenue model analysis
Variable and fixed cost estimation
Contribution margin calculation
Break-even calculation
ROI projection
Payback period calculation
Financial risk analysis
Confidence score generation

Tool used:

Calculator Tool

Output includes:

Startup costs
Monthly operating costs
Revenue model
ROI projection
Break-even analysis
Payback period
Financial risks
Confidence score

Retry Agent

The Retry Agent is designed to improve reliability when outputs are uncertain or confidence is low.

Used when:

Confidence score is too low
High-risk signals are detected
Business feasibility is uncertain
Finance or technology outputs indicate major concerns

Purpose:

Re-analyze weak areas
Improve decision reliability
Add a second-pass review pattern similar to real business decision workflows

Synthesis Agent

The Synthesis Agent combines outputs from all agents into a final decision report.

Responsibilities:

Combine research, technology, finance, and regulatory outputs
Generate a final Go / No-Go recommendation
Assign an overall confidence score
Summarize opportunities and risks
Include regulatory context when available
Provide practical next steps

Final output includes:

Final verdict
Overall confidence score
Research summary
Technology summary
Finance summary
Regulatory compliance summary
Consolidated risk list
Final recommendation

Retrieval-Augmented Generation (RAG)

This project includes two RAG pipelines.

1. Market RAG

Market RAG supports the Research Agent when the detected domain matches the available market documents.

Current Market RAG documents include:

Grocery delivery market context
Student spending behavior
AI personalization in retail
Delivery logistics risks

Market RAG workflow:

market_docs/
        ↓
Text Chunking
        ↓
OpenAI Embeddings
        ↓
ChromaDB Vector Store
        ↓
Similarity Search
        ↓
Research Agent Context Injection

Market RAG helps the system provide grounded insights about:

Customer demand
Competitors
Student pricing sensitivity
Delivery logistics risks
AI personalization opportunities
Campus-first launch strategy

2. Regulatory RAG

Regulatory RAG supports the Synthesis Agent by retrieving state-specific compliance context.

Supported states:

New Jersey (NJ)
New York (NY)
Pennsylvania (PA)
Connecticut (CT)
Massachusetts (MA)

Regulatory RAG retrieves context from official business registration and government sources.

Examples of retrieved information:

NJ-REG business registration requirement
State business formation guidance
Tax registration guidance
SBA business registration guidance

Regulatory RAG workflow:

Official Regulatory Sources
        ↓
Document Loading
        ↓
Text Chunking
        ↓
OpenAI Embeddings
        ↓
ChromaDB Vector Store
        ↓
State-Aware Retrieval
        ↓
Synthesis Agent Compliance Summary

Domain-Aware Routing

A domain detection layer was added to prevent unrelated RAG retrieval and reduce hallucinations.

Supported detected domains:

Grocery
Education
Healthcare
Finance
General business

Example behavior:

Input: AI grocery delivery startup in New Jersey
Detected Domain: grocery
Action: Uses Market RAG + Regulatory RAG

Input: I want to start a preschool program
Detected Domain: education
Action: Skips grocery Market RAG and performs generic education-focused business analysis

Input: Enterprise safety platform in New Jersey
Detected Domain: general
Action: Uses generic business analysis and state regulatory retrieval if NJ is detected

This improves answer validity by preventing grocery-specific context from leaking into unrelated domains.

Features Implemented

Core Features

PDF input support
TXT input support
Manual text input support
GPT-5 nano integration
OpenAI API integration
Local LLM fallback-ready structure
LangGraph multi-agent execution
Shared typed state across agents
Final synthesis generation

Tool Usage

Market context tool for Research Agent
Calculator Tool for Finance Agent
State extraction tool for Regulatory RAG
Domain detection tool for domain-aware routing

RAG Features

Market RAG using ChromaDB
Regulatory RAG using ChromaDB
OpenAI embedding-based retrieval
Domain-specific retrieval control
State-aware regulatory retrieval
Retrieved context preview in Streamlit

Evaluation Features

Rubric-based scoring
Research Agent scoring
Technology Agent scoring
Finance Agent scoring
Synthesis scoring
Consistency check across multiple runs
Confidence validation
Structured output evaluation

Streamlit UI Features

Professional dashboard interface
PDF/TXT upload
Manual business idea input
Domain detection display
State detection display
Executive summary cards
Final report rendering
Agent-level output tabs
Research RAG context viewer
Regulatory context viewer
Rubric evaluation display
Consistency check button
Download final report as Word document

Project Structure

Multi_Agent_Busisys/
│
├── agents/
│   ├── research_agent.py
│   ├── tech_agent.py
│   ├── finance_agent.py
│   ├── retry_agent.py
│   ├── synthesis_agent.py
│   └── __init__.py
│
├── app/
│   ├── main.py
│   ├── graph.py
│   ├── state.py
│   ├── llm_client.py
│   └── __init__.py
│
├── evals/
│   ├── rubric_eval.py
│   ├── consistency_check.py
│   └── __init__.py
│
├── rag/
│   ├── domain_detector.py
│   ├── extract_state.py
│   ├── ingest.py
│   ├── retrieve.py
│   ├── market_ingest.py
│   ├── market_retrieve.py
│   └── __init__.py
│
├── tools/
│   ├── calculator.py
│   ├── web_search.py
│   └── __init__.py
│
├── market_docs/
│   ├── grocery_delivery_market.txt
│   ├── student_spending_behavior.txt
│   ├── ai_personalization_retail.txt
│   └── delivery_logistics_risks.txt
│
├── input/
│   └── sample_business_idea.txt
│
├── output/
│
├── chroma_market/
├── chroma_regulatory/
│
├── streamlit_app.py
├── requirements.txt
├── .env
├── .gitignore
└── README.md

Note: chroma_market/ and chroma_regulatory/ are generated local vector database folders. They may be excluded from GitHub if the project rebuilds them using ingestion scripts.

Installation

Step 1 — Create Virtual Environment

py -m venv venv

Activate it:

.\\venv\\Scripts\\activate

Alternative Python 3.12 setup:

py -3.12 -m venv .venv312
.\\.venv312\\Scripts\\activate

Step 2 — Install Requirements

pip install -r requirements.txt

If needed, install additional packages:

pip install langchain-chroma python-docx

Step 3 — Add Environment Variables

Create a .env file in the project root:

OPENAI_API_KEY=your_api_key_here
LLM_BACKEND=openai
OPENAI_MODEL=gpt-5-nano
OPENAI_EMBEDDING_MODEL=text-embedding-3-small

Build Vector Stores

Market RAG Ingestion

Run:

py -m rag.market_ingest

Expected output:

Ingested 4 documents.
Created XX chunks.
Saved market vector store to: ./chroma_market

Regulatory RAG Ingestion

Run:

py -m rag.ingest

Expected output:

Indexing regulatory chunks into ChromaDB...
Vectorstore saved to ./chroma_regulatory

Some government pages may return 403 errors. This is acceptable if other official sources load successfully.

Running the Project

Terminal Version

py -m app.main

When prompted, enter:

input/sample_business_idea.txt

or any custom PDF/TXT business idea file path.

Streamlit Web App Version

streamlit run streamlit_app.py

The Streamlit app allows users to:

Upload PDF/TXT files
Type business ideas manually
Run complete multi-agent analysis
View final reports
Inspect all agent outputs
See RAG context
Check rubric scores
Run consistency checks
Download the final report as a Word document

Example Inputs

Grocery Startup Example

I want to launch an AI-powered grocery delivery startup in New Jersey for college students that predicts weekly needs and auto-suggests smart purchases based on budget and eating habits.

Expected behavior:

Detected Domain: grocery
Detected State: NJ
Market RAG used
Regulatory RAG retrieved
Final report includes NJ registration guidance

Preschool Program Example

I want to start a preschool program in New Jersey.

Expected behavior:

Detected Domain: education
Grocery Market RAG skipped
Regulatory RAG retrieved if state is detected
Output focuses on childcare, parents, staffing, tuition, curriculum, and licensing concerns

Enterprise Safety Platform Example

I want to build an enterprise safety and incident management platform for companies in New Jersey.

Expected behavior:

Detected Domain: general
Generic business reasoning
Regulatory RAG retrieved if state is detected
Output focuses on enterprise safety, compliance, security, operations, and risk management

Evaluation Metrics

Rubric-Based Scoring

Each agent is scored based on required outputs.

Examples:

Research Agent → market, competitors, demand, risk
Technology Agent → architecture, backend, timeline, risk
Finance Agent → cost, ROI, break-even, payback
Synthesis Layer → verdict, confidence, recommendation, risk

Consistency Check

The system can run the same input multiple times and compare final outputs.

It evaluates:

Recommendation stability
Confidence score consistency
Repeated conclusion alignment
Overall report consistency

Example:

Consistency Score: 87.5%

Example Final Output

Final Verdict:
Proceed with a campus-first MVP pilot in New Jersey.

Overall Confidence Score:
68/100

Research Summary:
Strong student-focused opportunity with budget-aware grocery planning.

Technology Summary:
Feasible using microservices, recommendation systems, API gateway, and cloud deployment.

Finance Summary:
Requires careful validation of delivery costs, ROI, break-even volume, and payback period.

Regulatory Summary:
NJ-REG business registration should be completed before operating in New Jersey.

Final Recommendation:
Launch a lean pilot, measure retention and unit economics, then scale if validated.

Course Concepts Used

Course Topic	Implementation
Prompt Engineering	Agent prompt design and structured instructions
Zero-shot Prompting	General business idea analysis
Few-shot Style Guidance	Agent-specific formatting and schema control
Tool Use	Calculator Tool and Market Context Tool
Retrieval-Augmented Generation	Market RAG and Regulatory RAG
Embeddings	OpenAI embeddings for vector retrieval
Vector Database	ChromaDB for local retrieval
Multi-Agent Systems	Research, Tech, Finance, Retry, and Synthesis Agents
LangGraph	Workflow orchestration and state passing
Evaluation	Rubric scoring and consistency checks
Structured Outputs	JSON outputs from agents
UI Development	Streamlit interactive dashboard
Document Generation	Downloadable Word report

Strengths of the System

Uses specialized agents instead of one generic prompt
Combines LLM reasoning with tools and retrieval
Supports domain-aware routing to reduce hallucinations
Retrieves official regulatory context when state is detected
Produces structured and explainable outputs
Provides evaluation metrics for agent performance
Includes a professional Streamlit demo interface
Generates downloadable reports for business review

Known Limitations

Market RAG is currently strongest for grocery/student-delivery business ideas
Regulatory RAG focuses mainly on business registration and state compliance sources
Some regulatory pages may block scraping or return 403 errors
Financial projections are estimated and should be validated with real business data
Domain detection is keyword-based and can be expanded with a classifier
Non-supported states will skip regulatory retrieval
Domain-specific RAG datasets need to be added for more industries

Future Enhancements

Add domain-specific RAG for education, healthcare, fintech, and security
Add real-time web search APIs
Add live market-size estimation
Add more robust domain classifier
Add deeper regulatory retrieval for childcare, healthcare, and food-delivery licenses
Add investor pitch deck generation
Add automated Excel financial model generation
Deploy Streamlit app to cloud
Add authentication and saved report history
Add comparison mode for multiple business ideas

Authors

Sri Harsha Chinta
Kamakshi Padma Vritika Manadapaka
Siddharth Raju Vysyaraju
Sri Lasya Siripurapu

Final Note

This project demonstrates how modern AI agents can support real-world business strategy decisions using structured reasoning, tool calling, retrieval-augmented generation, domain-aware routing, evaluation systems, and an interactive presentation layer.

It is designed to reflect realistic enterprise decision-making workflows rather than simple single-prompt LLM outputs.

Name		Name	Last commit message	Last commit date
Latest commit History 8 Commits
agents		agents
app		app
chroma_market		chroma_market
chroma_regulatory		chroma_regulatory
evals		evals
input		input
market_docs		market_docs
rag		rag
tools		tools
.gitignore		.gitignore
README.md		README.md
requirements.txt		requirements.txt
streamlit_app.py		streamlit_app.py

Folders and files

Latest commit

History

Repository files navigation

AI Multi-Agent Business Decision System

Project Overview

Business Problem

System Architecture

High-Level Workflow

Agent Workflow

Agent 1 — Research Agent

Agent 2 — Technology Agent

Agent 3 — Finance Agent

Retry Agent

Synthesis Agent

Retrieval-Augmented Generation (RAG)

1. Market RAG

2. Regulatory RAG

Domain-Aware Routing

Features Implemented

Core Features

Tool Usage

RAG Features

Evaluation Features

Streamlit UI Features

Project Structure

Installation

Step 1 — Create Virtual Environment

Step 2 — Install Requirements

Step 3 — Add Environment Variables

Build Vector Stores

Market RAG Ingestion

Regulatory RAG Ingestion

Running the Project

Terminal Version

Streamlit Web App Version

Example Inputs

Grocery Startup Example

Preschool Program Example

Enterprise Safety Platform Example

Evaluation Metrics

Rubric-Based Scoring

Consistency Check

Example Final Output

Course Concepts Used

Strengths of the System

Known Limitations

Future Enhancements

Authors

Final Note

About

Topics

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors

Uh oh!

Languages

Packages