Skip to content
    AndrewBuilds
    All projects

    Cognitive / AI

    COGNOS

    An AI cognitive operating system designed to capture raw thoughts, classify them, and map connections between ideas.

    Built with Base44AI APIsNLPGraph Architecture
    Live app

    Problem

    Notes apps store text; they do not store thinking. Ideas arrive unformed, at bad times, and the connection between two of them is usually the valuable part. COGNOS was an attempt to build a capture surface that classifies a raw thought the moment it lands and then shows how it relates to everything captured before.

    What I built

    • Thought capture
    • AI classification
    • Idea graph mapping

    Built by Andrew

    Inside the product

    The main screens and what each one does.

    • Instant entry
      No taxonomy
      Timestamped
      Offline-tolerant

      Raw capture

      One input box, no folders, no decisions. The cost of capturing a thought has to be near zero or the thought is lost.

    • Type detection
      Auto tags
      Confidence
      Manual correction

      Automatic classification

      Each entry is typed — idea, question, task, observation — and tagged, so structure appears after capture instead of blocking it.

    • Semantic links
      Clusters
      Timeline view
      Search

      Idea graph

      Related thoughts are linked by meaning, and clusters surface themes you did not know you kept returning to.

    Tech stack

    • Capture

      Fast entryVoice-to-text inputTag inference

    • AI layer

      ClassificationEmbeddingsSimilarity linkingNLP cleanup

    • Graph

      Node/edge modelCluster detectionInteractive map

    • Platform

      Base44Persistent store

    Challenges & how I solved them

    • Classification that is useful, not noisy

      Over-eager tagging made everything look connected. I constrained the label set and added a confidence threshold, so low-certainty entries stay unclassified rather than polluting the graph.

    • Meaningful edges

      Keyword overlap linked unrelated notes. Switching to embedding similarity with a minimum distance produced far fewer but far better connections.

    • Graphs that stay readable

      Past a few hundred nodes the map turned into hairball. Rendering now collapses clusters by default and expands on demand.

    Results so far

    Status
    LiveUsable at cognos-app.base44.app.
    Pipeline
    Capture → classify → connectFully automatic.
    Link quality
    Embedding-basedReplaced keyword matching.

    Related work