I propose AI Incubator, an LLM-powered tool for startup idea creation and validation. The process of early software startups shares similarities with ML research workflows. In ML research, we start with idea generation, write experimental code, analyze results, and iterate. The findings are then compiled into a paper to share with the research community. Similarly, in software startups, founders brainstorm ideas, develop an MVP, deploy it, conduct user interviews, and iterate. They may also create a report to present the idea to potential investors.
I modified the AI-Scientist framework to simulate the role of an early-stage software startup founder. Although the current results are not yet practical for real-world use, this proof of concept demonstrates the potential of applying the AI-Scientist structure to the domain of startups.
I integrated the following models into the framework:
- Deepseek-chat (Deepseek V3)
- ChatGPT-4o-latest
- Gemini-2.0-flash-exp
After some experimentation, Deepseek-chat was chosen for most tasks due to its low cost and decent quality.
Instead of generating research ideas, AI Incubator generates startup ideas. Each idea includes the following structured outputs:
- Title
- Name
- Problem
- Solution
- Target audience
- MVP
The LLM evaluates each idea based on scalability, feasibility, and novelty.
I introduced a new experimental "template" for startup MVP creation and validation. This template consists of:
- Skeleton Code: A basic structure for a simple CLI application.
- Survey Definition: Examples of survey questions—either open-ended or multiple-choice.
- User Interview Simulation: A function defined outside the template conducts user interviews. This function takes the CLI app, survey questions, and a description of the target user demographic as inputs.
User interviews are simulated by separate AI agents assigned roles matching the target audience's demographics. These agents test the app and answer survey questions based on their simulated experience. The framework enables AI Incubator to iteratively improve and validate the MVP using survey results.
At the end of the process, AI Incubator generates a report with the following sections:
- Background
- Problem Statement
- MVP Implementation
- Validation Approach
- Results
- Conclusion
For citations, I developed website_search.py using LangChain, allowing the LLM to source information from websites instead of relying on research papers.
For simplicity and to focus on the proof of concept, I removed:
- Idea novelty checks
- Paper review functionality
AI Incubator demonstrates the potential of AI to autonomously perform some early-stage software startup tasks. As foundational models continue to improve, this framework could evolve to better reflect real-world scenarios, such as supporting more sophisticated MVPs like fully-featured SaaS products.
One limitation of the current approach is that MVP testing and user interviews are simulated using AI agents instead of real users. However, AI agents proved very effective in identifying bugs. This framework of AI-based automated testing could provide a cost-effective solution for frequent product testing prior to deployment.