Kosmos

Our project, Kosmos, rethinks human-AI interaction by shifting away from the standard chatbox interface and into a spatial system for exploration. Instead of asking users to fully articulate what they want in a prompt, the interface lets them collect, arrange, and connect pieces of information like an idea bulletin board. Users can pin ideas, drop in references, and build a visual network of context through direct interaction. This makes the experience more attuned to how people actually think, which is often non-linear and driven by curiosity before clarity.
The core idea is simple: people often know what they are interested in before they know how to describe it. Our prototype helps users explore that gap between interest and expression. It surfaces connections between scattered inputs, reveals patterns across resources, and supports “rabbit hole” style discovery. This process does not force users into a rigid prompt response loop. By mapping thought visually instead of line by line, the system makes it easier to understand the bigger picture, revisit earlier branches, and reorient the direction of inquiry at any moment.
The working prototype achieves stability through a modular node-based interaction model. Each pinned item becomes part of a larger contextual system. As users add, remove, or rearrange inputs, the ambient interface updates the relationships and output structure consistently through the session, thanks to Redis as our database. The experience stays responsive and coherent even as the idea space changes. This creates a stable but flexible environment for exploration, where every interaction contributes to a shared context rather than starting over from scratch.
We specifically leverage Claude’s strengths in reasoning and large-scale context handling to interpret the full board as a connected system, not just isolated prompts. This allows the AI to infer intent, detect themes, and generate context-aware expansions. In this way, the UI is not just a wrapper around AI. The UI and the AI reasoning model work together as one simultaneous, generative experience.
This approach intentionally shatters common UI conventions in AI products. Instead of a single input field, a single conversation thread, and a single answer, we introduce a playful and spatial interaction pattern inspired by pin boards, threads, and idea mapping. Users can manifest their thought process via arrangement, proximity, and connection. This makes the interaction feel more human and less transactional. The result is a more intuitive workflow for exploration, research, and sense-making, especially in early-stage thinking when ideas are still forming.
Technologies, frameworks, and libraries used
Use this section and replace the placeholders with your exact stack.
Frontend framework: [Swift/SwiftUI]
Used to build a fast, component based interface for the interactive board, node system, and responsive layout.

UI rendering and interaction layer: [Swift/SwiftUI]
Enabled the visual pin board experience, draggable nodes, connection lines, and playful spatial interactions that move beyond a standard chat window.

State management: [Redis]
Managed shared board state, node relationships, and live updates as users add, remove, or rearrange inputs.

AI integration: [Claude API via Anthropic SDK, Websockets.]
Powered the reasoning layer that interprets the board as a whole, infers user intent, and generates context aware outputs from interconnected resources.

Backend or API layer: [Node.js ]
Handled secure AI requests, context packaging, and response orchestration between the interface and Claude.

Data persistence: [Redis]
Supported saving board state, restoring sessions, and preserving evolving idea maps over time.

Styling and responsive design: [Swift/SwiftUI]
Helped create a fluid, responsive UI that keeps the board usable across screen sizes while preserving visual clarity.

Why this stack supports the theme

Swift/SwiftUI made it possible to build an interface where intelligence is embedded in interaction itself. The frontend and visual interaction libraries support a responsive, spatial workflow for human exploration, while Claude via Node.js provides the reasoning engine that interprets relationships across the entire board. Together, they enable a new model of human and AI collaboration where users think with AI through arrangement, connection, and context, not just prompts.