DealFlow AI

DealFlow AI is a multi-agent AI system that performs institutional-grade investment due diligence in minutes instead of days. Built on CrewAI v1.0 with three specialized agents — Researcher, Financial Analyst, and Report Writer — it processes a startup pitch deck PDF and produces a comprehensive investment memo covering market validation, competitive analysis, financial projections, unit economics, and key risks. The system orchestrates Qwen2.5-72B-Instruct via HuggingFace Serverless Inference, with Serper.dev for real-time web search and ScrapeWebsiteTool for source verification. Each agent uses CrewAI's tool-calling primitives to gather, validate, and synthesize findings autonomously. The Researcher extracts claims and validates them against external market data. The Financial Analyst models projections, unit economics, and generates charts. The Report Writer synthesizes everything into a professional investment memo formatted for VC partners. Architecture is AMD MI300X-ready: the inference layer uses an OpenAI-compatible endpoint pattern, allowing drop-in deployment to AMD hardware. For this submission, the AMD Fallback Protocol activated at T+60h when MI300X credits did not arrive, switching to HuggingFace infrastructure while preserving the AMD-optimized architecture as the production path. This convertes a vendor limitation into an architectural strength: the system is portable across inference providers without code changes. Target users: small VC funds, angel investors, accelerators (Y Combinator, Techstars), and family offices. The cost reduction is significant — what traditionally requires $5,000-15,000 USD per startup in analyst time becomes $5-50 USD in compute costs per analysis. Built for the AMD Developer Hackathon 2026 by Peter-code-bot under Pedro Soriano Vergara at SORV Capital and Hackathon Squad Anáhuac Puebla.