Pixel Detective Platform

A local-first AI media search engine that brings semantic discovery, clustering, and curation to large personal archives. Built as a microservices platform with GPU acceleration.

FastAPI Services Qdrant Vector Search CUDA + cuML
Pixel Detective architecture

System Architecture

The platform is decoupled into a Next.js frontend, ingestion orchestration, ML inference, a vector database, and a GPU-accelerated UMAP service.

Frontend

Next.js 15, Chakra UI, React Query, Zustand, hydration-safe components.

Backend Services

Ingestion, ML inference, and GPU UMAP services coordinated via FastAPI.

Vector Intelligence

Qdrant stores embeddings and metadata for hybrid semantic search.

Performance

Dynamic batch sizing, FP16 models, and CUDA acceleration for throughput.

Core systems

Search Pipeline

Text and image queries embed via CLIP, then rank results in Qdrant.

Latent Space Explorer

UMAP projections, clustering, lasso selection, and multi-layer visualization.

Curation & Duplicates

Near-duplicate detection, archival workflows, and collection merging.

Streaming UMAP

Chunked processing with progress tracking for 10k+ images.

Feature visuals

Latent space explorer

Latent Space Explorer

Interactive scatter plots, clustering controls, and selection pipelines.

Curation workflows

Curation Suite

Duplicate analysis, archive safeguards, and collection-level actions.

Performance signals

Latency < 200ms Backend API response targets sustained.
UMAP < 2s for 500 points Interactive visualizations without blocking UI.
10x throughput GPU-optimized inference and batch autosizing.