Inspiration
Growing up alongside my family's business, I often helped prepare project proposals manually. That usually meant me spending hours on reading client requirements in a worn-down notebook, searching through old project files, copying relevant sections and rewriting everything into a new document.
Even today, sales teams and agencies and family businesses face the same problem. Valuable information is scattered across discovery call notes, past proposals, and project archives. Creating a high-quality proposal requires significant research and writing effort, which slows down the sales cycle and reduces productivity.
When I saw the Google Cloud Rapid Agent Hackathon and MongoDB MCP track, I realized this was the perfect opportunity to build the solution I wished had existed years ago: an AI agent that could automatically transform client conversations into professional business proposals while directly pulling information from an organization's historical knowledge base.
What it does
DealPilot is an autonomous AI sales proposal agent that converts discovery call transcripts into tailored business proposals.
Users simply paste a discovery call transcript, and the agent executes a transparent three-step workflow:
- Analyze extracts key business requirements from the transcript, including industry, company size, pain points, budget, timeline and objectives. Asks for any missing information.
- Research searches MongoDB Atlas for similar historical projects and relevant past experience from the company.
- Generate uses Gemini to create a professional 800–1200 word proposal customized to the client's needs.
The generated proposal can be viewed instantly, or exported as a PDF, and is automatically stored in MongoDB Atlas for future reuse.
Over time, DealPilot becomes smarter because every generated proposal expands the organization's knowledge base, creating a continuously improving proposal-generation system.
How we built it
DealPilot was built using Google's Agent Development Kit (ADK), Gemini, and MongoDB Atlas. The frontend is deployed on Vercel, while backend and agents are deployed on Google Cloud Run.
Agent Architecture
We designed DealPilot as a multi-step workflow:
- Analyze → extracts structured requirements from unstructured transcripts.
- Research → retrieves relevant historical projects from MongoDB.
- Generate → creates the final proposal using Gemini.
This modular design makes the workflow transparent, explainable, and easy to extend.
MongoDB MCP Integration
A custom MCP server powers DealPilot's knowledge retrieval layer.
Key MCP tools include:
search_similar_projects— finds relevant historical projects by industry.find_documents— queries MongoDB collections.save_proposal/insert_document— stores generated proposals.list_recent_proposals— retrieves recent outputs.aggregate_documents— provides analytics across proposal data.
MongoDB Atlas stores:
- projects collection containing seeded case studies, outcomes, budgets, and descriptions.
- proposals collection containing generated proposals, requirements and timestamps.
AI Proposal Generation
Gemini powers both information extraction and proposal generation, enabling DealPilot to transform unstructured conversations into professional sales documents with minimal human effort.
Additional Features
- Sample transcript loader
- Proposal history storage
- PDF export using the Anvil API
- Transparent step-by-step workflow visualization
- Reusable organizational knowledge base
Challenges we ran into
One of our biggest challenges was creating reliable communication between the AI agents and the MongoDB knowledge base.
Discovery call transcripts are highly unstructured, and extracting meaningful business requirements consistently required careful prompt engineering and workflow design.
Another challenge was ensuring that historical project retrieval actually improved proposal quality. We experimented with different retrieval approaches before settling on an industry-focused similarity search strategy that produced relevant examples without overwhelming the generator.
Building a smooth end-to-end workflow—from transcript ingestion to database retrieval, proposal generation, PDF export, and proposal persistence—also required coordinating multiple services and tools.
Accomplishments that we're proud of
- Successfully built a complete multi-step workflow using Google ADK.
- Integrated MongoDB through a custom MCP server instead of treating the database as simple storage.
- Created a reusable organizational memory system where every proposal improves future proposal generation.
- Implemented automated proposal generation from raw discovery call transcripts.
- Added PDF export functionality for real-world usability.
- Built a working end-to-end prototype during the hackathon that demonstrates a practical business use case for AI agents.
What we learned
This project taught us that AI agents become significantly more valuable when connected to organizational knowledge.
We learned how MCP can transform a database into an active tool that agents can query, analyze, and update during decision-making processes.
We also gained hands-on experience designing multi-agent architectures with Google ADK and learned how breaking complex tasks into specialized agents produces more reliable results than using a single monolithic prompt.
Most importantly, we learned that retrieval and memory are critical components for creating AI systems that deliver business value beyond simple text generation.
What's next for DealPilot
We see DealPilot evolving into a complete AI sales copilot.
Future plans include:
- CRM integrations (HubSpot, Salesforce, Zoho)
- Automated proposal personalization using company websites and public data
- Proposal scoring and win-probability prediction
- RAG-powered retrieval using vector embeddings
- Team collaboration and approval workflows
- Proposal versioning and analytics dashboards
- Support for contracts, statements of work, and pricing documents
- Continuous learning from accepted and rejected proposals
Our vision is to help sales teams spend less time writing proposals and more time building relationships and closing deals.
Log in or sign up for Devpost to join the conversation.