Project Catalog
AI-Powered Political File Management System
Technology Stack
PostgreSQL + pgvector
SQLAlchemy, Alembic
Anthropic Claude
Tesseract OCR
Vector Embeddings
Celery Workers
Cloud Storage
Rate Limiting
The Challenge
During the 2024 election cycle, designers were spending 1-5 minutes per file search in our Dropbox system. With hundreds of daily file requests, this inefficiency was costing the team hours of time weekly. While initially focused on search efficiency, deeper analysis revealed the core issue: account managers and designers needed a way to discover creative inspiration from past political content. Dropbox only searches vectorized words and titles, requiring users to recall specific keywords or file names rather than enabling semantic discovery of campaign themes, messaging strategies, or visual elements.
The Solution
I designed and built a full-stack AI-powered catalog system that automatically analyzes, tags, and semantically understands political design files. The application integrates with existing Dropbox workflows while providing intelligent search capabilities powered by vector embeddings and multi-modal AI analysis.
Key Capabilities
- • AI-powered content analysis and tagging
- • OCR text extraction from images
- • Semantic search with natural language queries
- • Custom political taxonomy mapping
- • Real-time processing with Celery workers
Technical Features
- • Multi-modal LLM integration (GPT-4, Claude)
- • Vector search with pgvector extension
- • Distributed task processing
- • Multi-layer caching architecture
- • RESTful API with FastAPI
Technical Architecture
System Design
- FastAPI Backend: High-performance async API with automatic OpenAPI documentation
- PostgreSQL + pgvector: Relational database with vector similarity search capabilities
- Celery Workers: Distributed task queue for background AI processing
- Redis Cache: High-performance caching layer for search results
AI Pipeline
- Document Ingestion: Multi-format file processing with preview generation
- OCR Extraction: Tesseract-powered text extraction from images
- AI Analysis: Multi-modal LLM analysis for content understanding
- Vector Embeddings: Semantic search with OpenAI embeddings
Technical Challenges Solved
Monolithic → Modular Architecture
Refactored from a monolithic Flask application to a modular FastAPI architecture with clear separation of concerns. Implemented service-oriented design with dedicated modules for AI processing, search, taxonomy management, and document handling.
Advanced Database Design
Architected a scalable PostgreSQL schema with pgvector extension for high-performance vector similarity search. Implemented complex many-to-many relationships for document-taxonomy mappings with optimized indexing strategies.
AI Prompt Engineering
Developed sophisticated prompt templates for political content analysis, achieving high accuracy in extracting campaign themes, candidate information, and visual elements. Implemented fallback strategies across multiple LLM providers.
Production Deployment
Containerized the entire application stack with Docker, implementing comprehensive security measures, rate limiting, and monitoring. Designed for cloud deployment with horizontal scaling capabilities.
KPIs
KPI Definitions
Design Relevance & Appropriateness
- • Compare recent designs to pre-system designs
- • Track client approval rates or revision requests
- • How well designs match campaign messaging/themes
Creative Diversity & Innovation
- • Are designers using more diverse design patterns?
- • Frequency of drawing from past successful campaigns
- • Using elements from different candidates/themes creatively
Design Process Quality
- • Fewer revisions needed when designers start with better inspiration
- • Can designers better explain their creative choices?
- • Time to reach a solid design direction (not just any design)
Key Features
Intelligent Search
Natural language queries like "healthcare policy graphics" or "Biden campaign materials" return semantically relevant results using vector similarity search and custom political taxonomy.
Automated Tagging
AI-powered analysis extracts political themes, visual elements, candidate information, and campaign messaging automatically from uploaded documents.
Analytics Dashboard
Minimal admin interface with usage metrics, search patterns, processing status, and system performance monitoring for continuous optimization.
Real-time Processing
Background processing with Celery workers ensures fast upload response times while comprehensive AI analysis happens asynchronously.
Explore the Project
Project Impact & Learning
This project demonstrates my ability to identify real business problems, architect scalable solutions, and deliver measurable impact through full-stack development and AI integration. It showcases advanced technical skills in modern web development, database design, and machine learning implementation.