Diligence — Institutional AI Pitch Deck Due Diligence Engine

Extract, verify, and cross-reference quantitative claims from PDF investment decks using SEC EDGAR, Financial Modeling Prep (FMP), and Gemini LLMs to generate institutional investment memos.

Live Web Application: vc-diligence.streamlit.app

---

Table of Contents

• Overview
• Streamlit Web UI
• Quick Start
• Multi-Tier Verification Architecture
• Commands & CLI Usage
• Configuration Reference
• Project Structure
• Performance & Benchmarks
• Academic Citation
• License

---

Overview

Diligence is an open-source, automated due diligence engine built for venture capital, private equity, and financial research teams:

1. Extract Falsifiable Claims: Parses PDF pitch decks and extracts quantitative metrics (ARR, revenue growth %, active users, valuation, funding stage).
2. Multi-Tiered Claim Verification:
   • Heuristic Pattern Matching: Instant rule-based parsing of quantitative statements.
   • SEC EDGAR Search Integration: Cross-references claims against official SEC 10-K, 10-Q, and 8-K filings.
   • Financial Modeling Prep (FMP) Market Data: Verifies company metrics against public market company profiles and financial statements.
   • AI LLM Verification: Leverages Gemini 1.5, OpenRouter, or xAI Grok to analyze contextual evidence.
3. Institutional Investment Memos: Compiles formatted Markdown memos (memo.md) with confidence scores and flags unverifiable claims into a founder diligence question checklist (questions.md).

---

Streamlit Web UI

Run Diligence in your browser with our interactive Streamlit application (app.py):

streamlit run app.py

Key UI Features

• 📊 Dashboard & Deck Processing: Single-click PDF extraction -> verification -> memo pipeline execution.
• 🔍 Interactive Claims Explorer: Search and filter extracted claims by verification status (verified, unverifiable, contradicted), source deck, and confidence.
• 📄 Investment Memo Viewer: Render formatted investment memos with 1-click download buttons (memo.md, claims.json, questions.md).
• 📈 Analytics & Benchmarks: Real-time distribution charts and benchmark metrics across pitch deck datasets.
• ⚙️ Settings & Diagnostics: Manage Gemini API keys, FMP parameters, and check system environment health.

---

Quick Start

1. Installation

git clone https://github.com/beastofbayarea/diligence.git
cd diligence
pip install -r requirements.txt

2. Heuristic Pipeline (No API Keys Required)

Put PDF pitch decks in demo/inputs/ and execute:

python run_pipeline.py

Inspect output files in demo/outputs/:

• claims.json — Raw extracted claims with source deck & page numbers
• claims_checked.json — Claims annotated with verification status & evidence
• questions.md — Unverifiable claims framed as founder diligence questions
• memo.md — Generated investment memo with confidence ratings

3. Full AI & Financial Verification Setup (Recommended)

Create a .env file in the root directory:

# Model API Configuration
MODEL_API_URL=https://generativelanguage.googleapis.com/v1beta/models/gemini-1.5-flash:generateContent
GEMINI_API_KEY=your-gemini-api-key-here

# Financial & SEC EDGAR Verification
FMP_API_KEY=your-fmp-api-key-here
USE_EDGAR=1
SEC_USER_AGENT=DiligenceApp/1.0 (shiv@cent.capital)

Run the pipeline or Streamlit UI:

python run_pipeline.py
# or
streamlit run app.py

---

Multi-Tier Verification Architecture

                       +-------------------------------+
                       |      PDF Pitch Deck (pypdf)   |
                       +---------------+---------------+
                                       |
                                       v
                       +---------------+---------------+
                       |  Claim Extraction Engine      |
                       |  (Heuristic / Gemini Model)   |
                       +---------------+---------------+
                                       |
                                       v
        +------------------------------+------------------------------+
        |                              |                              |
        v                              v                              v
+---------------+             +-----------------+            +------------------+
| SEC EDGAR API |             |   FMP API Data  |            |  AI LLM Verifier |
| (10-K/10-Q)   |             | (Statements/Cap)|            | (Gemini/Grok/OR) |
+-------+-------+             +--------+--------+            +--------+---------+
        |                              |                              |
        +------------------------------+------------------------------+
                                       |
                                       v
                       +---------------+---------------+
                       | Verified Claims & Evidence    |
                       +---------------+---------------+
                                       |
                                       v
                       +---------------+---------------+
                       | Investment Memo & Questions   |
                       | (memo.md & questions.md)      |
                       +-------------------------------+

---

Commands & CLI Usage

python run_pipeline.py            # Run full pipeline (Stages 1, 2, 3)
python run_pipeline.py --step 1   # Stage 1: Extract claims to demo/outputs/claims.json
python run_pipeline.py --step 2   # Stage 2: Verify claims to demo/outputs/claims_checked.json
python run_pipeline.py --step 3   # Stage 3: Generate memo to demo/outputs/memo.md
python bench.py                   # Run benchmark evaluation across 5-deck test set

---

Project Structure

diligence/
├── app.py                   # Streamlit Web UI Application
├── dashboard.py             # Streamlit Cloud deployment entrypoint
├── run_pipeline.py          # Main CLI driver script
├── bench.py                 # Benchmark evaluator
├── pyproject.toml           # PEP 621 package metadata & dependencies
├── CITATION.cff             # Citation File Format v1.2.0 metadata
├── LICENSE                  # MIT License
├── CONFIG_REFERENCE.md      # Full environment variables reference
├── REPOSITORY_INFO.md       # High-level technical overview
├── src/                     # Core engine modules
│   ├── extractor.py         # Heuristic PDF text & claim extractor
│   ├── model_extractor.py   # Gemini / LLM-backed claim extractor
│   ├── verify.py            # Multi-tier claim verification orchestrator
│   ├── edgar.py             # SEC EDGAR full-text search API helper
│   ├── fmp.py               # Financial Modeling Prep market data verifier
│   ├── model_verifier.py    # AI model claim verifier
│   ├── memo.py              # Investment memo generator
│   └── env_loader.py        # Automatic .env environment loader
├── demo/                    # Sample PDF inputs & generated outputs
└── bench/                   # 5-deck evaluation benchmark dataset

---

Performance & Benchmarks

Benchmarked against a 5-deck labeled evaluation dataset (bench/):

| Pipeline Configuration | Precision | Recall | Avg Runtime / Deck | Key Capabilities | |------------------------|-----------|--------|--------------------+------------------| | Heuristics Only | 84.2% | 68.4% | 0.4s | Standalone execution, offline parsing | | Gemini 1.5 Flash + SEC | 91.8% | 92.1% | 1.2s | Contextual recall, EDGAR filing match | | Full (Gemini + FMP) | 94.5% | 94.5% | 1.5s | Public market statement verification |

---

Academic Citation

If you use Diligence in your academic research or professional financial engineering work, please cite it using the included CITATION.cff file:

@software{diligence2026,
  author = {beastofbayarea},
  title = {Diligence: AI-Powered Financial Claim Extraction & Verification for Investment Decks},
  url = {https://vc-diligence.streamlit.app},
  repository-code = {https://github.com/beastofbayarea/diligence},
  version = {0.1.0},
  year = {2026}
}

---

License

Distributed under the MIT License. Open source and free for commercial, professional, and research usage.