VIA

Think Miro meets NotebookLM. A visual SECOND BRAIN where you chat, drop URLs, or files and watch AI build an interactive knowledge graph - cards, timelines, & insights that evolve as you keep learning

Comment

Inspiration

  • In 2021 I caught severe COVID and spent over a month in the hospital with ~75% of my lungs affected. Physically I recovered, but something subtle changed in how my mind held onto everyday experiences.
  • I started noticing that conversations, ideas, or things I had studied would fade completely after 3–4 weeks if I didn’t write them down. My brain would feel “blank” on topics I knew I had just gone through.
  • Interestingly, when I looked at my older written notes, the knowledge came back almost instantly; I could talk about those topics normally, as if the notes were “keys” that unlocked dormant memories.
  • People who went through severe COVID have reported similar “long COVID” symptoms: memory retention issues and difficulty recalling recent information. At the same time, everyone is drowning in content, videos, articles, lectures, that they can’t meaningfully retain.
  • Via Canvas is my attempt to build the tool I wish I had: a visual, AI-powered memory scaffold where everything I read, watch, or think about becomes a connected map I can revisit, grow, and be gently reminded of over time.

What it does

  • Via Canvas is like Miro meets NotebookLM: an infinite, colorful canvas where every idea, link, or lecture becomes a “card” on a living knowledge graph instead of disappearing into linear chat or static pages.
  • You talk to an AI side panel, paste URLs, or drop notes; the system auto-creates structured cards (notes, todos, links, videos, reminders) and places them intelligently on the canvas with connections.
  • A “Grow” action on any card asks the AI to expand that concept; creating child cards for subtopics, prerequisites, examples, or follow-up questions, turning single thoughts into rich branches of understanding.
  • A Timeline view lets you see what you learned and captured day by day, turning your knowledge into a visual memory diary instead of a flat note list.
  • The goal is simple: make it effortless for someone with a fragile or overloaded memory to retain, revisit, and grow what they’ve already invested time into learning.

How we built it

  • Frontend: React + TypeScript with ReactFlow powering the infinite canvas, custom card components for five card types, and Zustand for state management, undo/redo, and hierarchy operations (parent–child movement, collapse/expand, auto-layout).
  • Backend: An Express.js API that manages canvases, nodes, connections, and snapshots in PostgreSQL. This gives us a robust graph-like model with tags, card metadata, and auto-layout-friendly coordinates.
  • AI layer: A Python FastAPI service orchestrates “agents” (via NVIDIA NIM models) that can read the canvas, extract content from URLs, and decide what cards to create or how to “Grow” a concept, streaming responses back over SSE.
  • Canvas Intelligence: We implemented multiple auto-layout algorithms (tree, force-directed, circular), circular child arrangement around a parent, and synchronized movement of entire branches to keep the canvas readable as it grows.
  • All of this is wrapped in a dark, modern UI with vibrant card colors, rich interactions, and keyboard shortcuts so the experience feels like a serious tool, not just a prototype.

Challenges we ran into

  • Mapping chat to structure: It’s easy for AI to answer questions; it’s much harder to turn an ongoing conversation into a stable graph of cards, each with the right type, parent, and position, without overwhelming the user.
  • Auto-layout at scale: Making the canvas look clean with 10 cards is trivial; making it understandable with hundreds of cards, branches, and cross-links required careful layout algorithms, collision avoidance, and hierarchy management.
  • “Helpful but not bossy” AI: If the agent creates too many cards or connections, the canvas becomes noise. If it does too little, the user feels like they might as well be using a normal note app. Tuning that balance is non-trivial.
  • Temporal view design: Representing time in a way that is both functional (for review) and intuitive (as a “memory lane” of cards) forced us to design a separate timeline mode that still feels connected to the main canvas.
  • On a personal level, building a “memory prosthesis” while dealing with my own post-COVID memory issues was emotionally heavy—but also a strong motivator to keep the UX honest and practical.

Accomplishments that we’re proud of

  • We built a fully working infinite canvas with five distinct card types, tagging, auto-layout, intelligent child arrangement, collapse/expand, and undo/redo—all backed by a real database and API.
  • The Conversational Capture → Canvas loop is live: you can drop a URL or idea into chat, and see meaningful cards appear in the correct place with connections, not just plain text notes.
  • The “Grow” mechanic works end-to-end: right-click a concept and watch the AI generate structured child cards (e.g., prerequisites, subtopics, questions) and rearrange the local graph around it.
  • We implemented a timeline/chronological view that lets you visually replay what you learned and captured over days and weeks—a first step toward a true visual memory diary.
  • Most importantly, I already use Via Canvas for my own deep learning lectures, research notes, and “things to remember”—and it genuinely helps me recall and talk about topics I would otherwise forget.

What we learned

  • Chats are a terrible long-term memory format. Threads are great for interaction but terrible for resurfacing and recombining knowledge. A spatial, card-based graph works much better for revisiting and explaining what you know.
  • Giving users just enough structure (card types, tags, parent–child relationships) plus AI assistance is more powerful than either a blank whiteboard or a rigid note hierarchy.
  • Temporal context matters: seeing when you learned something—and what else you were learning near that time—adds its own kind of memory cue that pure text search can’t provide.
  • Architecturally, we learned how important it is to keep AI as a tool-using agent with access to canvas operations, not a black box that just returns text. That separation made the system more debuggable and extensible.
  • Building for a concrete personal constraint (post-COVID memory issues) forced us to design for real cognitive relief, not just “AI for the sake of AI.”

What’s next for VIA

  • Memory-first features: spaced-repetition style review modes, “remind me what I knew about X a month ago,” and gentle prompts to revisit stale branches of your knowledge graph.
  • Smarter agents: dedicated “Gap Finder,” “Teacher,” and “Research Synthesizer” agents that operate on your canvas to suggest what to learn next, challenge your understanding, or generate explanations and quizzes.
  • Deeper temporal views: richer timeline and “memory lane” experiences; seeing how entire domains (like “Deep Learning” or “Cybersecurity”) grew on your canvas across months and life phases.
  • Collaboration & sharing: shared canvases for study groups or teams, plus the ability to publish read-only “knowledge maps” so others can inherit or remix your understanding of a topic.
  • Longer term, I want Via Canvas to be not just a tool for “power users,” but a cognitive safety net for anyone dealing with long COVID, burnout, or simple information overload; a place where their daily learning doesn’t silently evaporate.

Built With

Share this project:

Updates