Inspiration: I am actually a student in India, and we used to have buildathons and hackathons. After attending some of these competitions in school, I realised that We spend 80% of our time in a project in Ideation, Looking at the maket, observing competitiors, feasibility etc. How about we automate these taks and make them faster and more predicatble. That's where it hit me to make genesis. Then 4 days before the submission data, I get an email from devpost, and then i challenge myself to make this project in 4 days.
What it does: It asks the user for a prompt on a feild on which the AI should brainstorm ideas. The user enters the feilds, and has an option to attach more files for reference. A startup idea is then given by the AI. The user gets 3 options after that, 1) Validate idea, 2) Deep VC analysis and 3) Unicorn prerdictor. Option 1 further validates the idea. Option 2 Stimulates a board room meeting with 3 different angentic AI personas, which are The brutalist Realist VC who criticizes the idea, The Visionary who defendes the idea and finally the Analyst who judges both the sides views and gives a final judgement. After that the user can generate 3 year predicted revenue model which can be taken to google sheets API
How we built it:Genesis is architected as a stateful, agentic AI platform designed to orchestrate multi-step reasoning, artifact generation, and persistent execution workflows. At the core of the system is a multi-agent debate loop powered by Gemini, where specialized agents with distinct system prompts (e.g., investor, market analyst, technical reviewer) operate within a controlled reasoning framework to perform adversarial evaluation, constraint checking, and consensus formation over a startup idea.
We used Python with the flask framework for the backend with a firebase database. The frontend was done in HTML CSS. Javascript was also used. We use Custom Google Search API for finding real time market data which we then process and feed to our gemini model. We have a fallback of using [link]ttps://hn.algolia.com/,[link]https://news.ycombinator.com and other sources if the API doesnt work.
The application backend is built as a modular service layer that manages session state, agent coordination, and output routing. Firebase is used as a persistence and synchronization layer, storing founding sessions, intermediate reasoning states, and generated artifacts, enabling continuity across interactions and providing an auditable decision trail.
Genesis integrates natively with Google Workspace APIs to convert AI reasoning into executable business outputs. A financial modeling service uses Gemini to construct structured revenue assumptions and projection logic, which is then programmatically written to Google Sheets using the Sheets API. Similarly, a presentation generation pipeline uses the Google Slides API to assemble investor-ready decks from structured narrative and visual components, ensuring consistency with validated strategy and financial data.
To bridge strategic reasoning with product design, Genesis includes a prototype simulation engine. Gemini generates valid Flutter (Dart) UI code based on the finalized concept, adhering to Material 3 design principles. This code is parsed and visualized within the application as a Figma-like interactive preview, allowing real-time inspection of layout hierarchy, component structure, and user flows without requiring a Flutter runtime or external tooling.
The system employs a hierarchical compute strategy, selectively invoking fast inference paths for exploratory analysis and deeper reasoning passes for strategic decision gates. This approach optimizes latency while preserving high-quality outputs for critical steps.
Overall, Genesis functions as an end-to-end autonomous execution system, combining agentic reasoning, real-time data grounding, artifact generation, and persistent state management to move beyond conversational AI and into deterministic, workflow-driven startup execution.
Challenges we ran into it: Due to lack of time and exams on the head, we faced time managemnet issues. There were some bugs that would take me hours to fix taking a lot of my time. less time meant lesser time to add more features we had planned to leading to more buggier features
Accomplishments that we're proud of:
1) Built a stateful multi-agent system that performs adversarial VC-style idea evaluation.
2) Generated real, editable financial models and pitch decks via Google Sheets and Slides APIs.
3) Implemented persistent founding sessions with Firebase for auditability and iteration.
4) Created a prototype simulator that generates Flutter UI code with an in-app visual preview.
5) Delivered an end-to-end autonomous execution workflow from idea to product artifacts.
What we learned: 1) Structured, adversarial reasoning produces better decisions than single-agent idea generation.
2) Turning AI outputs into real artifacts (Sheets, Slides, code) dramatically increases practical value.
3) Persistence and state are essential for moving beyond stateless chat into true execution workflows.
4) Tight API integration and scoped prompts matter more than model size for reliable results.
5) Shipping a focused, end-to-end system beats over-polishing individual components.
Log in or sign up for Devpost to join the conversation.