Diligent 🔍

AI-Powered M&A Due Diligence Assistant

Diligent assists deal teams by extracting structured data from financial documents, performing semantic search across datarooms, and providing AI-generated insights with full citations. Built for analysts who need fast document review with transparent, verifiable sources.

⚠️ Important: This tool assists analysts—it does not replace human judgment or comprehensive due diligence. Every answer must be verified against the cited source documents. This is an analytical aid, not a decision-making system.

---

🎬 Demo

Note: Demo video includes all features as of October 5th, 2025

---

� Use Cases We Cover

What This Tool Does:

1. 10-Q/10-K Report Analysis

   • Extracts 39 financial fields (revenue, EBITDA, debt, margins, cash flow, etc.)
   • Computes 5 financial ratios (D/E, net margin, ROE, debt coverage)
   • Retrieves specific sections with citations (document name + chunk location)
   • Investment Committee Memo Generation

2. CIM Review

   • Extracts market size, growth rates, customer concentration
   • Finds operational KPIs (CAC, LTV, churn, market share)
   • Searches across multiple sections with semantic matching

3. Contract Search & Extraction

   • Locates change-of-control clauses
   • Finds assignment provisions and key dates
   • Returns exact document location with page/chunk reference

4. Portfolio Dataroom Analysis

   • Batch processes multiple documents
   • Cross-document question answering
   • Tracks citations across entire dataroom

5. Risk Factor Identification

   • Extracts customer concentration, regulatory risks, litigation
   • Flags threshold violations (high debt, low margins)
   • Provides cited evidence for each finding

What This Tool Does NOT Do:

• ❌ Full document summarization (use structured data store for summaries)
• ❌ Period-over-period trend analysis (single snapshot only)
• ❌ Replace analyst judgment or comprehensive due diligence
• ❌ Legal contract analysis or advice

---

🏗️ Architecture

Data Flow

Upload sends PDFs to LandingAI ADE for parsing and structured field extraction (39 fields). Marker converts PDFs to markdown. Pathway normalizes extracted data and computes financial metrics. Text is chunked (1000 chars, 200 overlap) and indexed with semantic embeddings (sentence-transformers). Chat queries read from this indexed store: Gemini decomposes questions into sub-queries, hybrid RAG (70% semantic, 30% keyword) retrieves relevant chunks per sub-query, and Gemini synthesizes answers from structured data + retrieved spans. All citations include document name, chunk index, and sub-query attribution.

---

Tech Stack

Frontend:

• Next.js 14 + React 19 + TypeScript
• Tailwind CSS + shadcn/ui components
• Lucide Icons

Backend:

• FastAPI + SQLModel (SQLite)
• LandingAI ADE – PDF parsing + 39-field structured extraction
• Marker – PDF to Markdown conversion
• Pathway – Data normalization + financial ratio computation
• sentence-transformers – Semantic embeddings (all-MiniLM-L6-v2, 384-dim)
• Google Gemini AI (gemini-2.5-flash) – Query decomposition + answer synthesis
• ReportLab – PDF export

RAG System:

• Hybrid indexing (semantic + keyword)
• Multi-query decomposition
• Citation tracking with sub-query attribution
• JSON-based index storage (pathway_index/*.json)

---

Architecture & Data Flow

Upload Pipeline

User uploads PDF
    ↓
Save to ./data_rooms/{task_id}/{filename}
    ↓
Parse with Marker → Markdown text
    ↓
Extract with LandingAI ADE → 39 fields (JSON)
    ↓
Compute metrics with Pathway → 5 ratios
    ↓
Generate rule-based insights → ⚠️/✅ flags
    ↓
Index for RAG:
  • Chunk text (1000 chars, 200 overlap)
  • Generate embeddings (all-MiniLM-L6-v2)
  • Save to pathway_index/{task_id}_index.json
    ↓
Store in SQLite:
  • documents table: filename, filepath, markdown, extraction_json (39 fields)

Chat Pipeline (Multi-Query RAG)

User asks: "What are the financial risks?"
    ↓
Save question to chat_messages (status=pending)
    ↓
[BACKGROUND TASK]
    ↓
Load structured data from SQL (39 fields)
    ↓
Compute metrics with Pathway (5 ratios)
    ↓
Generate insights (threshold rules)
    ↓
Multi-Query RAG:
  1. Decompose query with Gemini → 3-5 sub-queries
     ["What is debt ratio?", "What are liquidity risks?", "What regulatory issues?"]
  
  2. For EACH sub-query:
     • Load RAG index (pathway_index/{task_id}_index.json)
     • Semantic search: cosine similarity (70%)
     • Keyword search: term frequency (30%)
     • Return top 5 chunks with metadata
  
  3. Collect citations:
     [{document: "10-Q.pdf", page: "Chunk 5", sub_query: "What is debt ratio?", ...}]
  
  4. Synthesize with Gemini:
     • Structured data (39 fields)
     • Computed metrics (5 ratios)
     • Rule insights
     • Retrieved chunks (grouped by sub-query)
     ↓
     Comprehensive answer in plain text
    ↓
Save to chat_messages:
  • content: answer text
  • reasoning_log: {sub_queries: [...], insights: [...]}
  • citations: [{document, page, sub_query, index}]
  • status: "done"
    ↓
Frontend displays:
  • Answer text
  • ▸ Analyzed via N queries (collapsible)
  • 📄 Sources: [doc1.pdf] [doc2.pdf] (pills)

---

Models & APIs Used

Service	Purpose	Cost
LandingAI ADE	PDF → Markdown + structured field extraction	Pay-per-document
Google Gemini AI	Natural language summary generation	Free tier: 15 req/min
Pathway	Data normalization and metric computation	Free (local library)

---

Installation

Prerequisites

• Node.js 18+
• Python 3.10+
• API keys: Google Gemini and LandingAI

Backend Setup

cd backend
python -m venv venv

# Windows
.\venv\Scripts\activate
# macOS/Linux
source venv/bin/activate

pip install -r requirements.txt

Create backend/.env:

#Enter your Gemini API key
GEMINI_API_KEY=your_gemini_api_key

#Enter your Landing AI API key
LANDINGAI_API_KEY=your_landingai_api_key

Run backend:

uvicorn main:app --reload --port 8000

Backend: http://localhost:8000 | API docs: http://localhost:8000/docs

---

Frontend Setup

cd frontend
pnpm install  # or npm install
pnpm dev      # or npm run dev

Frontend: http://localhost:3000

---

How to Use

1. Create Dataroom – Click "New Dataroom" in sidebar
2. Upload PDF – Must contain at least Revenue, Debt, and Equity fields
3. View Insights – Red flags popup shows leverage/liquidity warnings
4. Ask Questions – Type in chat (e.g., "What's the debt-to-equity ratio?")
5. Export Memo – Click "View Summary" → "Export to PDF"

---

Project Structure

backend/
  ├── main.py              # FastAPI app
  ├── database.py          # SQLite setup
  ├── models.py            # Task, Document, ChatMessage, Memo schemas
  ├── routes/
  │   ├── tasks.py         # Dataroom CRUD
  │   ├── documents.py     # Upload → ADE → Pathway pipeline
  │   ├── chat.py          # Gemini chat with structured metrics
  │   └── memo.py          # PDF export
  └── services/
      ├── landing_ai.py    # ADE API client
      ├── pathway_client.py # Data normalization + ratio computation
      ├── finance_logic.py  # Rule-based insight generation
      └── gemini_client.py  # Gemini API wrapper

frontend/
  ├── app/                 # Next.js pages
  ├── components/          # React components
  │   ├── chat-interface.tsx
  │   ├── dataroom-sidebar.tsx
  │   └── upload-modal.tsx
  └── lib/api.ts           # API client

License

MIT License