About the Project — Stellar
Inspiration
Stellar started around the beginning of this year as a personal experiment to build a research assistant. The original idea was simple: create an agent that could search the web, analyze information, and present structured results with minimal human input. I used the Tavily, OpenAI, and DeepSeek APIs to design a three-step workflow consisting of search, analysis, and output generation.
At first, the system only displayed results in the terminal. I later improved it to automatically generate PDF reports instead of plain text output. When I demonstrated this version to my teachers, they suggested turning it into a website so others could use it more easily. That feedback became the starting point for transforming Stellar from a script into a platform.
Development Journey
While rebuilding Stellar as a web application, I experimented with different language models and found that the Gemini 3 API produced longer and more consistent responses. Because of this, I migrated the core research pipeline to Gemini and turned the research module into the first dedicated agent, which I named Spectrum.
A major shift happened during a data analysis workshop at my college. Most participants were manually working through large Excel sheets, and I realized Stellar could automate the entire process. I updated the system to split large spreadsheets into smaller parts, summarize each section, and combine the results into a single dashboard. This became the second agent, called Cosmos, focused on data analysis and reporting.
Seeing platforms like Replit, Manus, and Gemini Build allow users to generate and deploy applications automatically inspired the next stage. I wanted Stellar to move beyond research and analytics and become capable of building software. This led to the creation of Nebula, a coding agent that plans applications, generates frontend and backend code, and verifies its own output.
The biggest limitation at that point was deployment. Stellar could generate code but had no way to run it in a real environment. One of my friends suggested integrating Docker, and that changed the entire project. With Docker support, Stellar could execute code in more than ten languages, create isolated containers, and host applications on a VPS. This functionality became a new agent named Forge, responsible for provisioning containers and deploying projects under a Stellar subdomain.
What I Learned
Building Stellar taught me far more than working with language models. I learned how to design multi-agent AI systems, manage different APIs, automate document generation, and process large datasets efficiently. The project also introduced me to Docker, container orchestration, and the practical challenges of deploying AI-generated software.
More importantly, I understood that an AI product is not just about model responses but about infrastructure, usability, and solving real problems for users.
Challenges
The project involved several technical difficulties. Different models produced inconsistent outputs, which made it hard to maintain a stable pipeline. Processing large Excel files required careful memory management. AI-generated code often failed in real environments, and connecting an LLM system with Docker securely took many iterations. Automating deployment on a VPS without breaking isolation was another major challenge.
Each of these problems forced me to rethink the architecture and gradually shaped Stellar into a more reliable platform.
Current State
Stellar is now a multi-agent system with four main components:
Spectrum handles research and report generation. Cosmos focuses on data analysis and dashboard creation. Nebula generates and verifies application code. Forge manages container provisioning and deployment.
What began as a terminal-based research script has evolved into a platform that can analyze data, generate software, and deploy applications globally.
Stellar uses the Gemini 3 API as the core reasoning model for research synthesis, spreadsheet summarization, and code generation across the Spectrum, Cosmos, and Nebula agents.
Log in or sign up for Devpost to join the conversation.