VentureForge

VentureForge is a hierarchical multi-agent system that automates startup idea discovery and validation. Built for the AMD AI Hackathon (Track 1: AI Agents & Agentic Workflows), it mines real user pain points from Hacker News, Product Hunt, and YouTube Data API, clusters similar complaints using LLM-based grouping, and generates startup ideas evaluated through Paul Graham-inspired binary rubrics. The system employs six specialized agents: an Orchestrator (supervisor), Pain Point Miner (scrapes HN, Product Hunt, and YouTube comments; clusters complaints), Idea Generator (creates startup concepts), Scorer (evaluates via 8 binary yes/no checks), Pitch Writer (produces investor-ready briefs), and Critic (adversarial review with 5 core checks). A reflection loop allows up to 3 revisions before final approval. Key innovation: all subjective evaluation uses binary yes/no checks instead of 0-1 scores, making results reproducible and auditable. The Scorer rubric covers feasibility, demand, and novelty (8 checks), while the Critic enforces evidence-backed claims with source URLs to prevent hallucinations. Built with LangGraph orchestration, Pydantic v2 validation, and Gradio UI. Production runs on AMD Instinct MI300X via vLLM + ROCm with Qwen/Qwen3.6-35B-A3B, featuring token optimization (one-at-a-time generation, compressed prompts) to work within the 2048 token limit.