This project was developed as part of a technical round assessment.
It is an AI-powered due diligence pipeline that takes a startup's pitch deck (PDF/PPT), performs web research, validates key claims, and generates a structured investment report for early-stage investors.
Note: Developed my first agent! And first ever multi-agent orchestrated pipeline. Pulled an all-nighter for this because it was too interesting, spent a lot of time navigating through a couple challenges with agent prompting, 202 site blockage errors (which I did not realize was due to iteratively testing my web scrapers) and I could NOT find any startup pitch decks (yes even with Google search hacks).
To build a Python app/script that:
- Ingests pitch decks in PDF/PPT format
- Extracts key information and claims
- Conducts external research beyond the deck
- Validates startup claims using multi-agent reasoning
- Outputs a well-structured Markdown report for an Investment Manager
Investment managers or analysts evaluating early-stage startups.
The system uses three coordinated AI agents, each with distinct roles:
| Agent | Description |
|---|---|
| Orchestrator Agent | Analyzes the deck, extracts company data, identifies gaps, and generates a research plan |
| Research Agent | Conducts parallel DuckDuckGo searches, scrapes content (via Selenium/Requests), and summarizes findings |
| Validator Agent | Fact-checks and verifies startup claims using structured evidence reasoning |
pip install requirements.txt
python pipeline.py <.pptx or .pdf file>
Input: Pitch deck (PDF or PPTX)
↓
Orchestrator → Extracts company info and creates research plan
↓
Research Agent → Executes searches, scrapes evidence
↓
Validator Agent → Validates claims using evidence
↓
Output: Markdown investment report
├── *_deck_analysis.json
├── *_research_results.json
├── *_validation_report.json
└── *_FINAL_REPORT.md # Main deliverable
- Add Slack bot, easy to use for Investment Manager
- Respond with markdown report in thread
- Support for more file formats
- Cached research results to avoid re-scraping
- Configurable validation thresholds
- Export to PDF/DOCX for sharing