Automated Deal Flow & Teaser Generation Pipeline

An end-to-end AI pipeline that converts unstructured company data into investment-grade M&A teaser decks.

This is a deterministic, reproducible pipeline designed to automate the early-stage investment analysis workflow typically performed by junior investment banking and private equity teams.

It ingests raw company data (PDFs, Excel models, Markdown briefs), enriches it with targeted public-web intelligence, extracts high-density, non-marketing investment facts using a rigorously constrained LLM agent, and programmatically renders:

• a native, editable PowerPoint teaser deck
• a citation audit document tracing data sources

No screenshots. No slide templates. No manual formatting.

---

🔍 Problem Statement

Early-stage M&A analysis suffers from three structural problems:

1. Unstructured inputs
   Company data arrives as PDFs, Excel sheets, one-pagers, and notes.

2. Low signal-to-noise summaries
   Generic AI summaries produce marketing fluff rather than investment facts.

3. Manual slide production
   Analysts spend hours formatting decks instead of thinking.

We solves this by enforcing:

• data density over prose
• strict schemas over free-form text
• code-driven slide construction

---

🧠 What This Pipeline Does (High Level)

1. Ingests private company data
   PDFs, Excel files, Markdown, and text files.

2. Augments missing context via public web search
   Focused queries for products, certifications, customers, and financials.

3. Builds a unified “truth context”
   Private + public data fused into a single analysis source.

4. Runs a highly constrained LLM extraction agent
   Outputs structured JSON only — no prose.

5. Renders investment slides programmatically
   Business overview, financials, KPIs, and thesis.

6. Generates a citation audit document
   Ensures traceability of claims.

---

🏗️ Architecture Overview

The system follows a GenAI-adapted ETL pipeline that converts unstructured company data into investment-grade outputs via structured LLM extraction and deterministic rendering.

---

📁 Project Structure

.
├── analyze.py # LLM agent & extraction logic
├── ingest.py # Data ingestion + web enrichment
├── ppt_engine.py # Programmatic PowerPoint renderer
├── generate_citations.py # Citation audit document generator
├── main.py # CLI entrypoint & pipeline orchestration
├── check_models.py # Gemini model availability checker
├── utils.py # Image download helper (Pexels)
├── requirements.txt # Python dependencies
├── examples/ # INPUT: files or data-pack folders
└── Final_Submissions/ # OUTPUT: PPT + citations

---

🔬 Pipeline Components (Deep Dive)

1️⃣ Ingestion Layer — ingest.py

Purpose:
Convert heterogeneous inputs into a single, analyzable text corpus.

Supported Inputs

• .pdf → parsed via LlamaParse (handles tables & scanned docs)
• .xlsx / .xls → flattened using pandas
• .md / .txt → read directly
• Folder-based “data packs” (multiple files per company)

Public Web Augmentation

• Uses Tavily Search
• Executes targeted queries for:
  • product capabilities
  • certifications & awards
  • revenue / geography indicators

Key Design Choice

The pipeline never assumes private data is complete.
Public data is always used to fill analytical gaps.

All extracted text is saved to: {CompanyName}_FULL_CONTEXT.txt for full transparency and debugging.

---

2️⃣ Extraction Agent — analyze.py

This is the core intelligence layer.

Model

• Google Gemini (gemini-2.5-flash or compatible)
• Chosen for:
  • very large context windows
  • reliable JSON compliance

Critical Constraints Enforced

• ❌ No marketing language
• ❌ No vague adjectives
• ❌ No “N/A”
• ✅ Numbers preferred over words
• ✅ Explicit inference with (est.) tagging
• ✅ Slide-ready sentences (≤20 words)

Output A rigid, predefined JSON schema covering:

• business overview
• infrastructure metrics
• product capabilities
• applications & certifications
• financial indicators
• investment thesis

Saved as: {CompanyName}_ANALYSIS.json

This file is the single source of truth for all downstream steps.

---

3️⃣ Presentation Engine — ppt_engine.py

Purpose:
Render investment-grade slides using code, not templates.

Key Characteristics

• Uses python-pptx
• Draws:
  • vector shapes
  • text boxes
  • metric tiles
  • native charts
• All text remains editable in PowerPoint

Slides Generated

1. Business Overview

   • company profile
   • infrastructure highlights
   • product & capability grid

2. Financial Performance

   • revenue & margin cards
   • revenue growth chart
   • operational KPIs

3. Investment Thesis

   • evidence-backed investment hooks

4. Legal Disclaimer

Design Philosophy

• No fragile XML hacks
• No version-specific PowerPoint features
• Clean, conservative “consulting-grade” layout

---

4️⃣ Citation Engine — generate_citations.py

Purpose:
Create an audit trail for extracted insights.

• Reads the citations field from analysis JSON
• Generates a .docx file listing data sources
• Intended for:
  • internal review
  • compliance checks
  • analyst validation

Output: Final_Submissions/{CompanyName}_Citations.docx

---

5️⃣ Orchestration Layer — main.py

This is the CLI controller.

Capabilities

• Detects whether input is:
  • a single file
  • a multi-file data pack folder
• Automatically derives company name
• Runs:
  1. analysis
  2. slide generation
  3. citation generation

Users interact via a simple numeric menu.

---

🛠️ Installation & Setup

Prerequisites

• Python 3.10+
• API keys for:
  • Google Gemini
  • Tavily Search
  • LlamaParse
  • Pexels (optional, for images)

---

1️⃣ Clone Repository

git clone https://github.com/neepun06/AI-ML-GC-RND.git
cd AI-ML-GC-RND

2️⃣ Create Virtual Environment

python -m venv venv source venv/bin/activate # macOS / Linux venv\Scripts\activate # Windows

3️⃣ Install Dependencies

pip install -r requirements.txt

4️⃣ Configure Environment Variables

Create a .env file in the project root:

GEMINI_API_KEY=your_gemini_key
TAVILY_API_KEY=your_tavily_key
LLAMA_CLOUD_API_KEY=your_llama_key
PEXELS_API_KEY=your_pexels_key   

▶️ How to Run

Step 1: Prepare Input

Option A — Single File

examples/
 └── Company-OnePager.md

Option B — Data Pack Folder

examples/
 └── Company_Data/
     ├── annual_report.pdf
     ├── financials.xlsx
     └── notes.md

Step 2: Run the pipeline

python main.py

Step 3: Select input

--- GENERATOR ---
Files and Data Pack Folders found:
[1] [FOLDER] Company_Data
[2] [FILE]   Company-OnePager.md

Select Item #: 1

Step 4: Retrieve Outputs

Final_Submissions/
 ├── Company_Teaser_Atomic.pptx
 └── Company_Citations.docx

⚙️ Design Decisions & Trade-offs

• Strict JSON schema was chosen over flexibility to guarantee slide safety.
• Estimation over N/A reflects real analyst behavior.
• Programmatic slides avoid template lock-in.
• Public web enrichment reduces dependency on perfect private data.
• Fail-soft design ensures partial data never breaks the pipeline.

🔐 Security & Data Handling

• API keys are excluded via .gitignore
• No data is transmitted except to configured APIs
• Generated outputs are local only

📄 License

Private project / hackathon submission. Not intended for public commercial redistribution.

📌 Final Note

This is a deterministic analytical system that treats LLMs as controlled extraction engines, not creative writers. This design choice is intentional.