Smart AI assistant for startups. From idea to public presentation – multi-agent system on Google Cloud ADK (Agent Development Kit) leads you to success. The project is inspired by the tasks of the hackathon "AI Agent Development Kit Hackathon with Google Cloud".
- 🔨 Architecture
- 🎯 Target Audience
- ⚒️ Technologies
- 📁 Structure
- ✨ Features
- 🔬 Testing
- 🏍️ How to run
- 📄 License & contribution
At the top level, the Coordinator/Dispatcher pattern is used for the venture_coordinator_agent, which redirects the user to the appropriate specialized agents. Crucially, the agents in this system don't merely execute tools; they leverage their LLM capabilities to interpret tool outputs, provide contextualized responses, and intelligently suggest next steps or delegate tasks to other specialized agents. This ensures a fluid, human-like interaction and guidance throughout the startup journey. The complete architecture is shown in the final diagram.
Venture Assist AI is designed for early-stage founders, solopreneurs, and startup teams who want to validate their ideas quickly, define their market strategy, and generate a compelling pitch — all without needing a full product or extensive team yet.
- Python v3.13.4, Uvicorn, Fastapi
- Node.js v22.16.0
- JavaScript, React, Vite
- Tailwindcss
- Docker
- Google Cloud
- Google Cloud SDK
- ADK (Agent Development Kit v1.1.1)
- Google AI Studio
- Generative AI on Google Cloud
- Google Gemini Pro/Flash
- Google Auth
- Google Workspace (Drive, Calendar, Meet, Slides)
- Secret Manager
- Artifact Registry
- Container Scanning
- Google Cloud Build, Google Cloud Run
- Firebase Hosting
venture-assist-ai/
│
├── backend/
│ ├── __init__.py # Initialize the package
│ ├── agent.py # Agent coordinator
│ ├── agents.py # Subagents
│ ├── config.py # Constants of models
│ ├── main.py # Entry point
│ ├── requirements.txt # Dependencies
│ ├── state.py # To store state
│ └── tools.py # Definitions of instruments
│
├── frontend/ # Style & UI design
│ └── ...
│
├── multi_tool_agent/ # Example a simple agent
│ └── ...
│
└── .env.example # Environment variables
A simple agent was developed using the Quickstart documentation.
Total 7 agents (1 coordinator & 6 subagents):
1️⃣ IdeaValidatorAgent -> get_validator
- Capabilities: Evaluates provided startup ideas.
- User Benefit: Provides intelligent analysis and detailed feedback on the idea.
2️⃣ MarketResearcherAgent -> get_research
- Capabilities: Conducts market research on a given topic.
- User Benefit: Generates a detailed report on market size, competitors, and trends.
3️⃣ PitchDeckGeneratorAgent -> get_pitch
- Capabilities: Creates a draft pitch deck for a startup.
- User Benefit: Forms compelling content for key presentation sections.
4️⃣ SummarySaverAgent
-> get_summary
-
Capabilities: Summarizes provided lengthy text.
-
User Benefit: Delivers a concise summary of notes on paper of a budding startup.
-
Improvement: Avoids unnecessary LLM calls for trivial cases using length check.
-> get_saver
-
Capabilities: Gets summary from memory and saves in Google Drive.
-
User Benefit: Shows info message if memory is empty.
5️⃣ LogoCreatorAgent -> get_logo
- Capabilities: Generates a logo concept for a startup idea.
- User Benefit: Suggests a creative logo idea with design and style descriptions.
- Improvement: Saves the logo description in Google Slides for future image creation.
6️⃣ MeetMakerAgent -> get_meeting
- Capabilities: Organizes a meeting with a specified participant.
- User Benefit: Extracts a date from the user input and writes the meeting to the calendar.
All agents not only confirms the validity of an idea but also provides contextual interpretation of the tool's output, suggesting further validation steps or delegating to another agent as needed.
The memory of the SummarySaverAgent is implemented via ToolContext, which provides an interface for interacting with the session state (tool_context.state). SessionService physically ensures the persistence of this state.
Thus, after summarization, the last_summary is stored in tool_context.state, and upon a subsequent save request, the agent utilizes this information either directly from arguments passed by the LLM or by accessing tool_context.state["last_summary"].
Implemented Google's basic authorization mechanism using two endpoints:
-
/auth/googleinitiates the Google authorization process and redirects the user to the consent page. -
/oauth2callbackaccepts the authorization code from Google, exchanges it for access and refresh tokens, and then stores them.
MeetMakerAgent uses the get_meeting tool, which calls extract_meeting_slots — a function powered by an LLM to propose meeting times.
Then Google services come into play:
-
Google Calendaris used to create the event. -
Google Meetis enabled automatically via the Calendar API when ads conference Data.
The LogoCreatorAgent uses the get_logo tool to use the Gemini model to generate a logo concept (including icon, colors, font, and mood). Then, interaction occurs with the Google Slides API: creates a new presentation, adds a blank slide to it, and inserts the generated logo concept into a text field on this slide. At the end, the function returns a confirmation of creation and a link to the created slide.
-> get_validator
Query:
Test the idea: a mobile app for finding nannies.
Query:
What do you think about the idea of the smart waste diversion system for households startup?
-> get_research
Query:
Conduct market research on the renewable energy sector.
Query:
Research the current state of the AI in healthcare market.
-> get_pitch
Query:
Generate a pitch deck for a web app that helps finding houses with AI.
Query:
Create a pitch deck for a AI-powered personalized learning platform, include sections for Problem, Solution, and Team.
-> get_summary
Query:
Summarize idea: An idea came to me, it seems a bit crazy, but the more I think about it, the more I like it. What if we make a smart mirror that helps track your mental state? Like, in the morning and in the evening you just go to the mirror, and it, looking at your facial expression, listening to your voice, intonation, evaluating your reactions, notices if you are burnt out, depressed or just tired. And it can gently suggest: "do a breathing practice", "try to rest a little", or even "it's time to talk to someone". Everything is local or encrypted. It's like a caring AI assistant, but not intrusive.
-> get_saver
Query:
Can you save it?
-> get_summary + get_saver
Query:
Summary and save: Listen, I have a cool idea. Why not make personalized tea based on DNA? Like, a person takes a simple test (like for genetics - saliva), plus fills out a questionnaire: how he sleeps, what he does, what flavors he likes. And then AI or an algorithm selects a tea blend for him - with the right herbs, vitamins, flavors and even the effect (calming, energy, recovery, etc.). You can do it as a subscription: every month a person gets "his" tea. It's like genetics + healthy lifestyle + a bit of a geek.
-> get_logo
Query:
Generate a logo concept for a startup that builds eco-friendly packaging solutions.
Query:
Create a logo for a fitness app for seniors.
-> get_meeting
Query:
Arrange a meeting with investor@gmail.com next week to discuss the startup.
Query:
Schedule a project discussion with alice@company.com for June 15th.
Query:
Organize an investor meeting with bob@example.com for tomorrow.
- Configure access:
Create .env (see example in .env.example).
- Authorize via Google OAuth 2.0:
gcloud auth application-default login
Be sure to check the all boxes.
- Run backend:
python -m venv .venv
.venv\Scripts\Activate.ps1 # for Windows PowerShell
pip install -r backend/requirements.txt
If want to testing agents via UI Google Agent:
adk web
For local server using frontend:
uvicorn backend.main:app --reload --port 8080
- Start frontend:
cd frontend
npm install
npm run dev
Project utilizes a robust cloud-based deployment strategy. The backend is deployed on Google Cloud services, leveraging:
-
Secret Manager: Securely storing sensitive data.
-
Artifact Registry: Serves as the repository for Docker images.
-
Container Scanning: Used for scanning Docker images for security.
-
Cloud Build: Automates the Docker image build process.
-
Cloud Run: Hosts the backend service as a serverless container.
The frontend is hosted on Firebase Hosting, providing a fast and reliable platform for the user interface.
This project is licensed under the MIT License. When using this software, please remember to provide appropriate attribution to the original author.
Note on the scope of the hackathon: For the purposes of this hackathon, the project is focused on demonstrating the functionality of the multi-agent system. Accordingly, some aspects, such as multi-user management and session persistence, have been simplified, assuming use by a single "test" user. The development of a full production version will include expansion of these capabilities.