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.
Category tags:
"This is an innovative system that automates startup idea discovery and validation. It addresses a real problem by mining real user pain points from multiple sources (Hacker News, Product Hunt, YouTube) and generating startup ideas evaluated through a structured, reproducible rubric. The hierarchical multi-agent architecture with 6 specialized agents and a reflection loop is impressive. The use of binary yes/no checks instead of scores for evaluation is a clever innovation for reproducibility and auditability. It runs on AMD MI300X with token optimization to fit the 2048 token limit. Application of Technology: 🚀🚀🚀🚀🚀 5 - Hierarchical multi-agent system with 6 specialized agents (Orchestrator, Pain Point Miner, Idea Generator, Scorer, Pitch Writer, Critic), LangGraph orchestration, Pydantic v2 validation, Gradio UI. Runs on AMD MI300X via vLLM + ROCm with Qwen/Qwen3.6-35B-A3B. Token optimization (one-at-a-time generation, compressed prompts). Presentation: 🚀🚀🚀🚀 4 - Clear explanation of the 6 agents and the process. Good problem statement. Links to GitHub and HuggingFace provided. Well-structured. Business Value: 🚀🚀🚀🚀🚀 5 - Massive market. Startup idea validation is a huge need. The systematic approach to mining pain points and evaluating ideas could help entrepreneurs and investors save significant time. The 8 binary rubric checks cover key aspects (feasibility, demand, novelty). Originality: 🚀🚀🚀🚀🚀 5 - Very original. Mining real pain points from multiple sources and generating startup ideas with a structured, reproducible evaluation system is unique. The binary yes/no check approach is a clever innovation for auditability."
Sanem Avcil
For Innovators & Creators