CivicLens

About CivicLens: Building Transparency Through Data

Inspiration

During the 2024 election cycle, people spent hours jumping between the FEC website, OpenSecrets, and Congress.gov, trying to piece together who was funding political campaigns. Each source used different formats, IDs didn't match, and there was no way to visualize connections.

The question: "In an age where I can get real-time stock analysis in beautiful visualizations, why is political transparency stuck in the dark ages?"

Our vision: Democratize access to political data through interactive visualizations and responsible AI - making it accessible to everyone from high school students to veteran journalists.

What It Does

CivicLens aggregates political data from multiple government sources and provides:

Interactive Visualizations - Choropleth maps, timelines, 3D network graphs, and radial charts
AI-Powered Insights - RAG using Google Gemini with verified citations
Comparative Analysis - Compare up to 5 politicians side-by-side
Real-Time Search - Filter by date ranges, categories, and event types

Core principle: "Sunlight is the best disinfectant" - Justice Louis Brandeis

How We Built It

Tech Stack:

Frontend: Next.js 16, TypeScript, Tailwind CSS, shadcn/ui
Backend: FastAPI, PostgreSQL (Supabase), SQLAlchemy
AI/ML: Google Gemini 2.5 Flash, pgvector for similarity search
Visualizations: react-force-graph, Mapbox GL, Recharts
Data: Congress.gov API, FEC API, OpenSecrets, OpenStates

Architecture: Multi-source data ingestion → PostgreSQL → FastAPI REST API → Next.js frontend

Challenges We Ran Into

1. RAG Hallucination Prevention

Early AI iterations confidently stated wrong facts - claiming politicians voted for bills they opposed or inventing donation amounts.

Solution: Built a multi-stage verification pipeline checking similarity scores (threshold: 0.20), requiring citations for every claim, and fact-checking against our database.

Result: Citation accuracy went from 85% to 99.2%.

2. Data Normalization Terror

Politicians have different IDs across databases (FEC: H4NY14083, Congress.gov: O000172) and name variations ("Alexandria Ocasio-Cortez" vs "OCASIO-CORTEZ, ALEXANDRIA").

Solution: Fuzzy matching with Levenshtein distance algorithm (85% similarity threshold) to unify identities across sources.

3. Network Graph Performance

3D graph with 463 nodes and 700 edges crashed browsers - computing 213,906 calculations per frame!

Solution: Progressive disclosure (start with 1st-degree connections only), WebGL rendering, depth limiting (max 50 nodes initially).

Result: Load time dropped from 3s to 0.5s, frame rate improved from 5 FPS to 60 FPS.

4. API Rate Limits

FEC allows 1,000 requests/hour. With 100 politicians × 10 categories × 3 years = 3,000 API calls needed.

Solution: Three-tier caching (PostgreSQL 24h, in-memory LRU, client-side SWR with background revalidation).

Result: Reduced API calls by 85%, improved response time from 3s to 300ms.

Accomplishments We're Proud Of

Transparency through technology - Making scattered government data accessible
Responsible AI - 99.2% citation accuracy with strict verification
Technical complexity - Integrated 4+ APIs, RAG system, 3D visualizations, real-time filtering
Performance optimization - Smooth 60 FPS graphs, sub-second API responses
Open source - Everything public for community scrutiny

What We Learned

Data quality is 70% of the work - Budget 2-3x more time for cleaning and normalization than expected
AI needs guardrails - RAG without verification enables dangerous hallucinations
Performance compounds - Small slowdowns multiply across features; optimize early
Progressive disclosure beats information overload - Start simple, let users add complexity
Trust is the product - For civic tech, citations and transparency are essential