FounderOS

Inspiration Every year, 9 out of 10 startups fail — and the #1 reason is "no market need." Founders pour months into building before discovering their idea doesn't hold up. We asked: what if an AI-powered advisory board could stress-test your startup idea in minutes, not months? We were inspired by how VCs and accelerators evaluate startups — multiple expert perspectives, live market data, financial modeling, and a final invest/pass verdict — and set out to democratize that entire process using autonomous AI agents.

What it does FounderOS takes a raw startup idea — just a sentence — and runs it through 10 specialized AI agents orchestrated with CrewAI. In minutes, you get:

Idea Clarification — structures your raw thought into a clear problem/solution/value prop Live Market Research — real competitors, segments, and trends pulled from the web with cited sources (via Tavily) Positioning & ICP — who your ideal customer is and how you're differentiated 4-Week MVP Roadmap — prioritized features and a build plan Landing Page Copy & Pricing — marketing-ready messaging and pricing tiers Investor Debate — a Bull investor argues the upside, a Skeptic attacks weaknesses, and a Moderator synthesizes both into a verdict Financial Model — unit economics snapshot GO / NO-GO Scorecard — a final recommendation with next experiments Everything streams to the frontend in real-time via SSE, and the results are stored as a competitive knowledge graph in Neo4j.

How we built it Backend: FastAPI + Python with Pydantic models for strict data validation, running on Uvicorn AI Agents: 10 specialized agents built with CrewAI + LangChain, powered by NVIDIA NIM (Kimi K2.5) via OpenAI-compatible API Live Research: Tavily API for real-time web search with citations Knowledge Graph: Neo4j to store and query competitive landscapes Frontend: Next.js 14 + React 18 + TypeScript + Tailwind CSS, with Lucide React for icons and react-force-graph-2d for graph visualization Real-time Streaming: SSE (Server-Sent Events) via sse-starlette for live agent progress Database: PostgreSQL (via Render) for persistence, with local JSON storage fallback PDF Reports: Auto-generated venture analysis reports Integration Adapters: Modular clients for Yutori, Senso, Modulate, and Numeric with graceful fallback stubs Challenges we ran into Agent coordination — getting 10 CrewAI agents to pass structured state sequentially without data loss; we solved it with a single StartupDossier Pydantic model that accumulates results across all agents Parsing LLM outputs to JSON — LLMs return free-form text; we built a parser that extracts JSON from markdown code blocks and handles edge cases Real-time streaming — wiring SSE from the FastAPI backend through each agent step to the Next.js frontend with proper event handling NVIDIA NIM integration — routing CrewAI/LangChain through NVIDIA's OpenAI-compatible endpoint for Kimi K2.5 required careful configuration of the base URL and model name Graph visualization — rendering Neo4j competitive landscape data as an interactive force-directed graph in the browser using react-force-graph-2d Accomplishments that we're proud of 10 autonomous agents working as a cohesive pipeline — not one LLM call, but a true multi-agent system with CrewAI Every market claim is cited with real URLs from Tavily — no hallucinated competitor data The investor debate is genuinely useful — the Bull vs. Skeptic synthesis surfaces real risks and opportunities that you'd miss doing it alone Live streaming UX — users watch each agent think in real-time, which keeps them engaged during the 3-5 minute pipeline Modular architecture — 6 integration adapters with stub fallbacks mean the system runs even without every API key What we learned Multi-agent orchestration with CrewAI is powerful but demands careful state management — a shared Pydantic data model is essential SSE streaming dramatically improves UX for long-running AI workflows — users stay engaged when they can watch progress Knowledge graphs (Neo4j) add a queryable dimension that flat JSON can't — competitive landscapes become explorable and visual NVIDIA NIM provides a seamless OpenAI-compatible gateway to powerful models like Kimi K2.5, making it easy to swap LLM providers Graceful degradation matters — stub adapters let the full system work in dev without every third-party API key configured What's next for FounderOS Parallel agent execution to cut run time from 5 minutes to under 2 Industry-specific agent templates (fintech, healthtech, SaaS) with fine-tuned prompts Team collaboration — share reports, comment, and iterate on ideas together Auto-generated pitch decks from the analysis output Deeper graph analytics — path-to-market recommendations from Neo4j competitive data Benchmarking — compare your idea against similar funded startups to predict fundraiseability

Built With

• amazon-web-services
• crewai
• docker
• fastapi
• fastino
• kimi-k2.5
• neo4j
• nvidia-nim
• postgresql
• python
• react
• render
• sse
• tailwind-css
• tavily
• vite