Live by Flash

Live agent research navigator for literature overload

Comment

Inspiration

Researchers today face a paradox: there is more knowledge available than ever before, but understanding the relationships between ideas has become increasingly difficult. Publication growth is accelerating with a doubling time of 17.3 years (Bornmann 2021). Decades ago, a researcher might read 20 core papers a year; today that number can exceed 2,000. Scientific papers, policy reports, and technical research are scattered across thousands of sources, and traditional tools treat them like isolated documents rather than connected systems of knowledge.

We were inspired by the question: What if research could be explored like a living world instead of a list of PDFs?

The idea behind Live by Flash is to turn collections of papers into interactive knowledge frontiers. Instead of searching documents individually, users can visualize clusters of ideas, explore relationships between competing research, and interact with an AI agent that understands the evolving landscape of a topic.

Our goal was to combine AI agents, generative visualization, and research analysis to create a new way to navigate complex domains such as climate resilience, healthcare, policy, or artificial intelligence itself.

⚡ Powered by Gemini

Live by Flash uses Google Gemini models to:

  • Understand and summarize research papers
  • Perform semantic clustering across documents
  • Power a conversational agent for frontier exploration
  • Generate contextual visual environments

This enables a multimodal research experience combining text, structure, and visual reasoning.

What it does

Live by Flash transforms collections of research papers and sources into AI-generated knowledge landscapes.

Users can add papers, links, or research sources to create a Frontier — a workspace that represents a specific research domain. Once sources are added, the system: Navigates full-text corpora, inferring relationship beyond citations Builds an interactive directed acyclic graph mapping claims, evidence, methods, and contradictions across the literature Clusters papers into related themes and debates Generates a visual research landscape Places clusters within that landscape Allows users to interact with a live AI agent to explore insights Flags weak evidence, surfacing conflicting results, and linking each conclusion to source excerpts

Instead of navigating papers individually, users explore an interactive research frontier where clusters represent key ideas such as: Nature-based solutions Policy and governance Data and modeling Infrastructure adaptation

The AI agent can answer questions like: Where does the research consensus exist? Which papers disagree? What strategies appear most effective? What gaps exist in the literature?

The result is a multimodal research interface where visual context, semantic clustering, and conversational AI work together. Keeping humans in the loop, the agent augments expert reasoning, automates repetitive navigation and evidence-tracing tasks, and leaves interpretation and final judgment to researchers.

How we built it

Live by Flash was built as a modern AI-driven research interface using a combination of frontend visualization and generative AI services.

Key components include: AI research ingestion When a user adds a paper or link, the system uses Gemini models to: Read and summarize research sources Extract themes and metadata Generate structured research objects

Semantic clustering The papers are then sent to Gemini for semantic clustering, which groups related research into conceptual clusters rather than relying on predefined categories.

Knowledge landscape generation Using generative image models, the system creates a contextual research environment based on the themes of the frontier. This visual landscape acts as a canvas where clusters can be explored. Interactive exploration

The interface supports multiple views: research landscape view, graph view, document inspection view

Users can also click clusters to inspect sources and interact with a Frontier Agent that answers questions about the research space.

Live agent interaction A conversational AI agent powered by Gemini allows users to ask natural-language questions about the research frontier and receive synthesized insights drawn from the clustered papers.

Built with: TypeScript, React, Vite, Gemini API, Google GenAI SDK, Google Maps Platform, Node.js,G Google Cloud -Run (planned deployment), Generative AI image models, Semantic clustering algorithms, and React Flow visualization

Challenges we ran into

Building Live by Flash involved several technical and design challenges.

AI rate limits and orchestration Because clustering and summarization require multiple AI calls, we had to carefully control when requests were made to avoid hitting API limits during development.

Representing research spatially One of the biggest challenges was designing a visual system that could represent abstract research relationships in an intuitive way. Traditional maps rely on geographic coordinates, but research ideas are conceptual rather than spatial.

We experimented with geographic mapping before evolving the design toward AI-generated knowledge landscapes that better reflect thematic relationships.

Integrating generative media with data structures Generating images that meaningfully represent research topics required combining structured data with generative prompts and building a pipeline that keeps the visuals synchronized with the clusters.

Balancing automation and user control AI can automatically analyze research, but we also needed to give users control over when analysis runs and how the frontier evolves.

Accomplishments that we're proud of

We are proud that Live by Flash demonstrates a new paradigm for navigating complex knowledge.

Some highlights include: Turning research papers into interactive knowledge frontiers Building an AI clustering system that identifies relationships between papers Top-tier visual design to enhance interpretation Memorable, high-fidelity graphics overlay the research landscape so key patterns are instantly recognizable without replacing expert thought. Cluster metaphors: when a group of papers converges on a key trend, an expressive, landmark appears to represent that theme. Integrating multiple forms of AI: text understanding, clustering, and generative imagery Most importantly, the project shows how AI can transform static research documents into explorable knowledge systems.

What we learned

Through this project we learned several important lessons about building AI-native interfaces.

First, AI is most powerful when it augments exploration rather than replacing it. Instead of generating answers in isolation, the system becomes much more useful when it helps users navigate complex information spaces.

Second, multimodal interfaces — combining visualization, generative imagery, and conversational agents — can dramatically improve how people understand complex research topics.

Finally, building AI systems requires careful orchestration between models, data structures, and user experience. The hardest problems are often not model accuracy but designing interactions that make AI insights understandable and trustworthy.

What's next for Live by Flash

Live by Flash is just the beginning of a broader vision for AI-powered research exploration.

Next steps include:

  1. Voice interaction with the Frontier Agent using Gemini Live
  2. Real-time web research to add new sources automatically
  3. Collaborative frontiers where teams explore research together
  4. Advanced disagreement mapping between competing papers
  5. Integration with research databases and academic repositories

Ultimately, we envision Live by Flash becoming a platform where researchers, policymakers, and innovators can explore complex knowledge spaces as living systems of ideas.

Built With

  • cloud-run
  • cloud-run-(planned-deployment)
  • gemini-api
  • generative-ai-image-models
  • google-genai-sdk
  • google-maps-platform
  • node.js
  • react
  • semantic-clustering-algorithms
  • typescript
  • vite
Share this project:

Updates