A Multi-Agent Orchestration Framework for Venture Capital Due Diligence

In one sentence

Summary

The paper addresses the inefficiencies and hallucination risks inherent in manual venture capital due diligence by proposing a fully automated, event-driven multi-agent orchestration framework. The system decomposes the research process into specialized sub-tasks handled by distinct AI agents, combining large language models with real-time web retrieval to synthesize unstructured data into structured investment intelligence. This approach aims to reduce human error, cognitive bias, and delays while ensuring that financial figures are grounded in auditable, official sources rather than generated estimates. The methodology employs a Directed Acyclic Graph (DAG) architecture implemented on the n8n low-code platform. The pipeline begins with data intake, followed by market and competitive intelligence modules that query the Perplexity Sonar Deep Research API for real-time sector sizing and competitor analysis. A core technical contribution is a programmatic extraction pipeline that reverse-engineers the frontend-to-backend communication of the Greek Business Registry (Γ.E.MH.) to retrieve official financial PDFs, which are then parsed using a layout-aware OCR extractor. The system includes a structural fallback mechanism that queries third-party databases like Crunchbase or Dealroom if registry data is unavailable, or explicitly flags data absence to prevent hallucination. The framework produces structured HTML reports containing company overviews, market intelligence, competitive landscapes, and financial summaries with preserved source citations. Key findings highlight the system's ability to automate end-to-end corporate research, delivering actionable insights such as attractiveness scores and 30-to-180-day recommendations. However, the system is limited by its reliance on the Greek registry for high-quality financial data, dependency on external commercial APIs, and non-deterministic LLM outputs. The authors note that generalizability across geographies and sectors requires further empirical evaluation, and suggest future work includes integrating international registries and self-hosted alternatives to reduce operational costs.

Research problem

Manual venture capital due diligence is prone to human error, cognitive bias, and delays, while standard LLMs suffer from knowledge cut-offs and hallucination risks when generating financial figures. The paper addresses the need for an automated, reliable system that can synthesize fragmented, unstructured data from multiple sources into structured investment intelligence without fabricating data.

Core contributions

• A modular, multi-agent architecture that automates end-to-end corporate research from a single analyst trigger to a structured HTML report.
• A programmatic financial extraction pipeline that reverse-engineers the Greek Business Registry (Γ.E.MH.) endpoints to retrieve and parse official financial filings.
• A structural fallback mechanism that explicitly flags data absence rather than generating unverified figures, directly targeting hallucination in financial contexts.
• Public availability of all workflow artifacts to support replication and further research in agentic financial workflows.

Method

• Event-driven orchestration on the n8n low-code platform using a Directed Acyclic Graph (DAG) of processing nodes.
• Specialized AI agents for data intake, market intelligence, financial extraction, and synthesis, each with distinct roles and tools.
• Reverse-engineering of Γ.E.MH. frontend-to-backend communication to query dynamic endpoints for official PDF filings.
• Layout-aware OCR extraction (using Marker) to parse financial statements and preserve source citations.
• Conditional routing to third-party databases (Crunchbase, Dealroom) or explicit 'Not Found' flags when registry data is absent.
• Integration with Perplexity Sonar API for real-time web search and market sizing.
• Automated generation of structured HTML reports delivered via Gmail API.

Results & findings

• The framework successfully automates the synthesis of unstructured data into structured investment intelligence, reducing manual effort.
• The reverse-engineered Γ.E.MH. pipeline allows access to official financial filings without administrative API credentials.
• The structural fallback mechanism effectively prevents hallucination by flagging data gaps instead of generating plausible but false figures.
• Generated reports include actionable outputs such as attractiveness scores and specific 30-to-180-day recommendations for funds and startups.
• The system demonstrates the viability of low-code, event-driven orchestration for complex financial workflows.

Data & experiments

• The system was deployed and tested within a single Greek venture capital fund (Uni.Fund) for operational due diligence.
• Data sources include the Greek Business Registry (Γ.E.MH.) for official financial filings, Perplexity Sonar API for real-time market intelligence, and third-party databases like Crunchbase and Dealroom.
• The pipeline processes unstructured web data, news, and scanned PDF documents to generate structured financial metrics and competitive analyses.
• Evaluation is primarily qualitative, focusing on the reliability, traceability, and structure of the generated reports rather than quantitative trading performance.

Limitations & caveats

• The financial extraction component is specific to the Greek Business Registry, limiting high-quality data to Greek-incorporated entities.
• The pipeline depends on external commercial services (Perplexity, LLM provider, PDF extractor), creating risks from pricing changes and rate limits.
• Generalizability across investment stages, sectors, and geographies remains to be established through broader empirical evaluation.
• LLM inference is non-deterministic, leading to potential variability in synthesis outputs across repeated runs.
• The system's effectiveness relies on the quality and availability of data in third-party databases for non-Greek companies.

Curator notes

Discovery status: institutional_trading_focus_promoted. Source links: LLM venture capital due diligence investment