DealFlow AI 🎯

Multi-agent M&A due diligence — powered by Band

DealFlow AI is a cross-framework multi-agent system that performs complete M&A due diligence in minutes. Six specialized AI agents collaborate through Band's shared environment — passing structured signals, coordinating tasks, and producing a final investment memo without any human in the loop.

Live demo: dealflow-agent-production-4bf4.up.railway.app

---

How it works

Submit a company name and any available documents. Six agents immediately spring into action inside a shared Band room, routing work to each other, posting structured signals, and handing off outputs until the investment memo is generated.

User submits company → Orchestrator kicks off pipeline
    ├── WebResearch    → scrapes market data → posts SIGNAL:market_research
    ├── DocumentParser → extracts financials from uploaded files → posts SIGNAL:parsed_documents
    │       ├── FinancialAnalyst → CAGR, burn rate, valuation range → SIGNAL:financial_analysis
    │       └── LegalRisk        → CoC clauses, IP risks, liability → SIGNAL:legal_risk
    └── Synthesis → reads all signals → writes investment memo PDF → SIGNAL:investment_memo

All coordination happens through Band — agents @mention each other, post structured JSON signals, and the Band room acts as the shared memory and message bus for the entire workflow.

---

The 6 agents

Agent	Character	Framework	Model	Role
Orchestrator	🎩 Conductor	OpenAI Agents SDK	GPT-4o	Kicks off the pipeline, routes tasks, escalates human review
WebResearch	🔍 Detective	LangGraph	GPT-4o	Wikipedia + DuckDuckGo market research, posts to Band directly
DocumentParser	📚 Librarian	OpenAI Agents SDK	GPT-4o-mini	Extracts financials, contracts, and cap table from uploaded PDFs
FinancialAnalyst	💼 Banker	OpenAI Agents SDK	GPT-4o-mini	Calculates CAGR, burn rate, runway, and valuation range
LegalRisk	⚖️ Judge	OpenAI Agents SDK	GPT-4o-mini	Flags change-of-control clauses, IP risks, and liability exposure
Synthesis	🔮 Wizard	OpenAI Agents SDK	GPT-4o	Synthesizes all signals into a final PDF investment memo

---

Band integration

Band is the central nervous system of DealFlow AI — not just a notification layer.

• Every agent listens on a shared Band room via the thenvoi SDK
• Agents post structured JSON signals (e.g. SIGNAL:financial_analysis) that downstream agents parse
• @mentions route work: @FinancialAnalyst only activates when it receives a message with its handle
• The Orchestrator reads incoming Band messages to decide what to trigger next
• WebResearch posts its signal directly via the Band REST API (bypassing the LLM) for reliability
• The Band room acts as a persistent shared log — Synthesis reads the full message history to compile its memo
• If LegalRisk flags a deal-breaker, the Orchestrator posts an escalation message to loop in a human reviewer

---

Tech stack

• Agent frameworks: OpenAI Agents SDK + LangGraph (cross-framework)
• Models: GPT-4o and GPT-4o-mini via AI/ML API
• Coordination: Band (thenvoi) — shared room, @mentions, structured signals
• API: FastAPI (Python)
• Frontend: Vanilla JS + CSS animations
• Deployment: Railway (Docker)

---

Running locally

Prerequisites

• Python 3.11+
• uv
• A Band account — get free Pro with code BANDHACK26 at band.ai/manage-billing
• AI/ML API key for GPT-4o access

Setup

git clone https://github.com/<your-username>/dealflow-agent
cd dealflow-agent

uv sync

cp .env.example .env
# Add your AIML_API_KEY and ORCHESTRATOR_API_KEY

cp agent_config.yaml.example agent_config.yaml
# Add your Band agent IDs and API keys for all 6 agents

agent_config.yaml

orchestrator:
  agent_id: "..."
  api_key: "..."
web_research:
  agent_id: "..."
  api_key: "..."
document_parser:
  agent_id: "..."
  api_key: "..."
financial_analyst:
  agent_id: "..."
  api_key: "..."
legal_risk:
  agent_id: "..."
  api_key: "..."
synthesis:
  agent_id: "..."
  api_key: "..."

Run

# Terminal 1 — start all 6 agents
python run_agents.py

# Terminal 2 — start the API + frontend
uvicorn api.main:app --reload --port 8000

Open http://localhost:8000 and submit a company.

---

API

# Submit a company for analysis
curl -X POST http://localhost:8000/analyze \
  -F "company_name=Stripe" \
  -F "notes=Series B SaaS, considering acquisition" \
  -F "files=@financials.pdf"

# Upload a document to an existing deal (saves under uploads/{deal_id}/, posts DOCUMENT_UPLOADED to Band)
curl -X POST "http://localhost:8000/deals/{deal_id}/upload" \
  -F "file=@./pitch.pdf"

# Check status
curl http://localhost:8000/deals/{deal_id}

# Download the investment memo PDF (most recently completed deal)
curl http://localhost:8000/memo/latest --output memo.pdf

# Deal Score + verdict JSON for the UI (same “latest” deal)
curl http://localhost:8000/memo/latest/summary

# When synthesis finishes, register the PDF and optional memo summary (JSON body)
curl -X POST "http://localhost:8000/deals/{deal_id}/complete" \
  -H "Content-Type: application/json" \
  -d '{"memo_summary":{"deal_score":72,"risks_flagged_count":3,"company_name":"Acme","recommendation":"conditional","confidence":"medium","executive_summary":"Summary text here."}}'

---

Project structure

dealflow-agent/
├── agents/
│   ├── orchestrator/      # GPT-4o, OpenAI Agents SDK
│   ├── web_research/      # GPT-4o, LangGraph
│   ├── document_parser/   # GPT-4o-mini, OpenAI Agents SDK
│   ├── financial_analyst/ # GPT-4o-mini, OpenAI Agents SDK
│   ├── legal_risk/        # GPT-4o-mini, OpenAI Agents SDK
│   └── synthesis/         # GPT-4o, OpenAI Agents SDK
├── api/
│   └── main.py            # FastAPI gateway
├── shared/
│   ├── prompts.py         # System prompts for all agents
│   └── models.py          # Shared Pydantic models
├── frontend/
│   ├── index.html         # Single-page UI
│   └── characters/        # Agent character images
├── Dockerfile
└── railway.json

---

Why multi-agent?

M&A due diligence is inherently parallel and multi-disciplinary. A single agent can't simultaneously hold deep financial modeling, legal contract review, market research, and narrative synthesis in context without degrading on all of them.

By splitting into specialists that communicate through Band:

• WebResearch and DocumentParser run concurrently — no waiting
• Each agent's context window is focused on one domain
• Outputs are structured and auditable — every signal is a JSON blob in the Band room log
• Human escalation is built in — LegalRisk can trigger a human review without breaking the pipeline

---

Built for the Band of Agents Hackathon · June 2026