Technical Summary: A Multi-Agent Orchestration Framework for Venture Capital Due Diligence

Problem Statement

Venture capital due diligence and portfolio monitoring currently rely on manual desk research to synthesize fragmented information from incompatible sources, including unstructured web data, real-time news, and official financial registries often locked behind administrative controls. This manual approach introduces human error, cognitive bias, and significant delays. While Large Language Models (LLMs) offer powerful natural language understanding, their direct application to financial due diligence faces three structural limitations:

1. Knowledge Cutoffs: Training data limitations render models blind to current market signals.
2. Hallucination: LLMs frequently generate plausible but factually incorrect information, a failure mode unacceptable for financial decision-making.
3. Lack of Autonomy: Base LLMs cannot autonomously query external databases, invoke APIs, or parse scanned financial documents without specialized frameworks.

Methodology

The authors propose a fully automated, event-driven multi-agent orchestration pipeline implemented on n8n, a low-code platform. The system operates as a Directed Acyclic Graph (DAG) of processing nodes, triggered by a single analyst input via a custom HTML form. The workflow is decomposed into four distinct stages:

1. Data Intake and Context Initialization

Upon trigger, a webhook captures the payload and maps the target company to a pre-scraped JSON database containing baseline attributes (founders, sector, registration numbers). An AI Context Agent synthesizes this into a compact profile to anchor downstream research.

2. Market and Competitive Intelligence

Five specialized agents utilize the Perplexity Sonar Deep Research API for real-time intelligence:

• AI Source Mapper: Identifies reliable sector-specific portals and aggregators.
• AI Sector & AI Competition: Run in parallel to extract market sizing, macro-trends, and categorize competitors (direct, adjacent, niche).
• AI News & AI Signals: Scan for strategic developments (partnerships, launches) and extract quantifiable investment signals.

3. Financial Data Retrieval and Fallback

This component addresses the challenge of accessing official financial filings, specifically from the Greek Business Registry (Γ.E.MH.).

• Reverse-Engineering: Instead of relying on an official API (which requires administrative credentials), the framework reverse-engineers the registry's frontend-to-backend communication to query dynamic endpoints.
• Document Processing: The system retrieves raw PDF documents (balance sheets, income statements) and classifies them into "Corporate Modifications" and "Financial Statements."
• Layout-Aware OCR: A specialized extractor (Marker) processes PDFs to preserve structural integrity, followed by summarization agents that distill metrics (Assets, Liabilities, Revenue, EBIT) with source citations.
• Structural Fallback: If a company lacks a valid Greek registry number (e.g., foreign incorporation), the pipeline routes to a Alternative Financials Agent. This agent queries third-party commercial databases (Crunchbase, Dealroom). Crucially, if no data is found, the system explicitly flags the entry as "Not Found" rather than generating an estimate.

4. Synthesis and Output Generation

Three synthesis agents aggregate the intelligence streams:

• AI Researcher: Creates a strategic research note highlighting developments and blind spots.
• AI Analyst: Drafts an Executive Summary, assigns attractiveness scores (market timing, product differentiation), and formulates 30-to-180-day recommendations for both the fund and the startup.
• AI Overall Company Info: Generates a public-facing summary for founders.
  The final output is a styled HTML report delivered via Gmail API, with every claim traceable to its specific intelligence source.

Key Contributions

The paper identifies three primary technical contributions:

1. Modular Multi-Agent Architecture: An end-to-end automation system that transforms a single analyst trigger into a structured, email-delivered report, lowering the engineering barrier for investment teams via low-code implementation.
2. Programmatic Financial Extraction Pipeline: A novel method for querying hidden registry endpoints, retrieving raw PDFs, and applying layout-aware OCR to structure financial data with preserved citations, specifically targeting the Greek Business Registry.
3. Structural Fallback Mechanism: A design pattern that explicitly routes to third-party databases or flags data absence when official registry data is unavailable. This directly mitigates hallucination by replacing unverified figures with auditable data gaps.

Results and System Behavior

The system successfully automates the synthesis of unstructured market signals and official regulatory filings. The output is a structured HTML report where the provenance of every claim is traceable to specific agents and data sources.

• Hallucination Mitigation: The system demonstrates its ability to handle missing data by rendering explicit "Not Found" flags in the Financial Summary block when registry data is absent, rather than fabricating figures.
• Epistemic Clarity: The report distinguishes between three states: registry-sourced figures (auditable), third-party commercial approximations, and explicit data gaps, preventing the conflation of these distinct data qualities.

Significance and Scope

The authors position this work as a step toward automating pre-investment due diligence with a reliability level previously requiring human analysts. The significance lies in the integration of dynamic web retrieval, reverse-engineered endpoint querying, and OCR-based document processing into a single pipeline.

The paper maintains a modest scope, acknowledging that:

• The financial extraction is currently specific to the Greek Business Registry, limiting coverage to Greek-incorporated entities.
• The system relies on external commercial services (Perplexity, LLM providers, PDF extractors), introducing dependencies on pricing and interface stability.
• The current deployment is validated within a single Greek venture capital fund, and generalizability across geographies and sectors requires further empirical evaluation.
• LLM inference remains non-deterministic, with potential variability in synthesis outputs that is not yet quantified.

Future work proposed by the authors includes integrating internal portfolio databases for contextual depth, expanding to international registries (e.g., UK Companies House), and replacing third-party PDF extraction with self-hosted open-source alternatives.