Developed by Fred McCullough
Interactive Three.js explorer for inspecting semantic search relationships.
This project turns embedding-based similarity across 8,406 publicly available Montgomery County business records into a browser-based visual surface. It supports concept-based discovery, neighborhood inspection, and map handoff without reducing the work to a flat directory.
The corrected LinkedIn announcement uses this same live screenshot set; the project proof is preserved directly below for readers who arrive through GitHub first.
I build systems that turn messy real-world data into inspectable workflows. This project demonstrates how embedding-based retrieval can be made visible and navigable, allowing users to walk a semantic neighborhood and carry a focused result into map context.
- Interactive 3D Navigation: Pan, zoom, and rotate through a dynamically generated semantic constellation.
- Concept-Based Discovery: Visual grouping and color-coding of semantically related data points (e.g., "coffee shops," "law firms") based on vector proximity.
- Guided Camera Choreography: Smooth camera movement that keeps selected records and nearest-neighbor trails understandable.
- High-Performance Rendering: Built on Three.js for smooth web-based 3D graphics, even with thousands of data points.
- Responsive HUD: A tailored heads-up display that provides real-time metadata for the focused semantic neighborhood.
- Three.js Core: Utilizes custom shaders and instanced rendering for optimal performance.
- Vector Mapping: Pre-computed semantic threads are loaded and visualized to represent data relationships.
- Dynamic Physics: Custom particle physics and glow effects keep the dense graph readable while preserving the sense of a live network.
- Semantic Backend: The
backend/directory contains the Python pipeline used to generate and serve embeddings and nearest-neighbor artifacts. - Local Model Cache Layer: The live Hostinger deployment includes a guarded local inference worker that precomputes cached "Deep trail note" artifacts for selected semantic trails. Public visitors only read cached artifacts through a read-only API path; cache misses fall back silently to deterministic guide copy instead of starting large-model generation.
For a deep dive into the engineering decisions and systems mindset behind this project, see the McCullough Digital systems page or follow the LinkedIn announcement.
| Artifact | What it shows |
|---|---|
index.html |
Full browser experience and interaction model |
semantic-demo.css |
Responsive UI, HUD styling, search/focus states, and motion polish |
backend/ |
Semantic artifact generation path behind the visualization |