Frontier

• 

  Initial page where you can input your starting vector (defining your research vibes)

• 

  The Gemini CLI agent automates file creation and experiments and manages the project structure and findings.md in real-time.

• 

  Findings.md which shows results and graphs from the experiment.

Frontier - Platform for vibe research so anyone can contribute to frontier science

Inspiration Science shouldn't be locked behind ivory towers. We believe that anyone with a strong intuition and a passion for a field should be able to contribute to human knowledge. We were inspired by the "vibes" and "inklings" people have about how the world works, those initial ideas that often get lost because the path to a formal hypothesis is daunting. We built Frontier to bridge that gap and turn every curious mind into a researcher.

What it does Frontier acts as an AI research co-scientist. You start by sharing your high-level thoughts or "inklings" about a research field. While you chat, Frontier builds a dynamic knowledge graph in real-time. Each node represents a specific concept, with many linking directly to academic papers.

The AI uses this graph to ask pointed, deep-diving questions that map out your intuition against existing literature. Once it identifies a genuine research gap, it helps you boil everything down into a concrete hypothesis. But it doesn't stop there. Frontier then hands off the hypothesis to a Gemini CLI agent. This agent performs targeted searches, writes the experimental code, executes it, and generates a findings.md. Finally, Frontier presents the results and drafts a professional abstract for you.

How we built it We leaned heavily into the Google ecosystem and leveraged the multimodal and coding capabilities of Gemini 3. We used Google AI Studio to prototype the initial conversational flow and knowledge graph logic. For the heavy lifting of the experimental phase, we integrated the Gemini CLI. In a bit of a "meta" move, we actually used the Gemini CLI to help us write the integration code for the CLI agent within the app itself. The core platform is built with TypeScript, while Python handles the data processing and experiment execution scripts. We deployed with Google Cloud Run.

Challenges we faced The biggest hurdle was ensuring the AI didn't just summarize existing info but actually pushed the user toward new territory. Mapping messy human intuition to a structured knowledge graph required complex prompt engineering and state management. Integrating the CLI agent to reliably write and run code based on a brand-new hypothesis was also a major technical challenge that required rigorous testing.

What we learned We learned that AI is most powerful when it acts as a mirror for human creativity. By visualizing the research field as a graph, we saw how quickly a vague idea can become a structured scientific inquiry. We also realized that "prompting" can be much more than just text. It can be a collaborative, multi-step journey from an inkling to a discovery.

What's next for Frontier We want to expand Frontier to support more complex simulation environments and integrate with open-access journals to automatically suggest even more relevant sources. Our goal is to make "doing science" as accessible as sending a text, helping the world solve problems faster through decentralized, AI empowered research. We also want to integrate Gemini Deep Research so it can verify its research direction truly is novel and can refine the hypothesis off of existing research.

Built With

• gemini
• typescript