Movement Language Discovery System

Approach

The system is built as a Next.js web installation (hosted on Vercel) backed by Firestore and Google Cloud Storage, with Google Video Intelligence handling automatic labeling. Its actual design work is the gap it deliberately leaves open: every submission shows machine-generated tags next to a blank "Community says" column, making the absence of cultural context visible rather than papering over it with a synthetic label. A password-gated reviewer dashboard turns that gap into a decision point — full clip playback, existing annotations, and a three-state action (Approve/Restrict/Reject) with a required reason field, so nothing enters the public record unread. The visual language reinforces the same tension: Fraunces for archival/editorial content, DM Mono explicitly for anything machine-generated, and a water-refraction motion motif standing in for how meaning distorts when it passes through a system not built to hold it. The companion system, The Repository, is the payoff — approved knowledge lives on a Raspberry Pi with no internet connection, no cloud vendor, and no third-party platform between the community and its own archive.

Getting There

• —Reviewed critiques of Laban Movement Analysis (LMA) as a taxonomy — including its origin in European concert dance and challenges to its claimed universality — before deciding to use it anyway, explicitly as a placeholder classification backbone rather than a claim of correctness
• —Made the schema-level decision to store physical description separate from cultural meaning, so the same movement can carry contested or multiple interpretations without the database forcing a single 'truth'
• —Confronted a sourcing problem directly: pulling cultural movement data from research sites or scraped web content would replicate the exact extraction-without-consent problem the project argues against, so used publicly self-documented sources and clearly-labelled synthetic data instead
• —Visitor records or uploads a movement video, required to name themselves and describe the movement's meaning before submitting
• —Video uploads to Google Cloud Storage via a Next.js server-side API route — chosen over Firebase Storage after discovering Video Intelligence requires gs:// URIs rather than HTTPS URLs, removing an unnecessary abstraction layer
• —Google Video Intelligence auto-tags the video with generic, scene-level labels; a Firestore record tracks status at each pipeline stage
• —Visitor receives a QR-coded permanent link to their submission, with an empty 'Community says' column left for the reviewer — the gap made visible rather than hidden
• —A human reviewer reads the submission, adds up to five cultural annotations, and decides Approve, Restrict, or Reject; every decision is logged and timestamped
• —Approved submissions are exported as markdown and ingested into The Repository — an offline RAG system (LangChain, ChromaDB, Llama 3.1 8B via Ollama) running on a Raspberry Pi, with zero cloud dependency
• —Deliberately kept the web pilot and the offline RAG device as separate systems with separate repos, rather than merging them via a Firebase-coupled path that was considered and rejected
• —Measured actual performance on the Pi rather than assuming it would be fast enough for a design-show demo