Technology
Indexing 330,000+ government entities takes real engineering: high-throughput data pipelines, geospatial search, AI-driven document and speech processing, and resilient infrastructure spanning origin, cloud, and edge. Here's what powers the platform.
Core Platform
Python↗
Primary language across every service and script
FastAPI↗
Async web framework serving all API endpoints and pages
Uvicorn↗
ASGI server running production workloads
Jinja2↗
Server-side HTML templating with i18n support, for the live apps and for every static site we publish
Starlette↗
ASGI toolkit for middleware, sessions, and request/response handling
asgi-lifespan↗
Programmatic ASGI lifespan management for background service startup and shutdown
Pydantic↗
Data validation and serialization for API models and configuration
asyncpg↗
High-performance async PostgreSQL driver for all database access
Babel↗
Internationalization engine powering 18-language locale support
Doppler↗
Secrets management injecting credentials across all services
tldextract↗
Accurate eTLD+1 extraction from URLs for referrer domain bucketing
Google Analytics Data API↗
GA4 Data API client powering weekly referrer digest reports
RapidFuzz↗
Fuzzy name matching: speakers, commenters, and roll-call spellings in meeting transcripts, and report titles across sources
nh3↗
HTML sanitization for rendered Markdown and inbound mail before either reaches a page
boto3↗
S3-compatible client that uploads the static sites and service-request media to Cloudflare R2
Database
PostgreSQL↗
Primary datastore for all entities, crawl data, and service requests
PostGIS↗
Spatial extension for geographic boundaries, area queries, and map rendering
pgvector↗
Vector similarity search for semantic document matching
Alembic↗
Versioned schema migrations: every database change is a reviewed revision, applied the same way to the origin database and the cloud tier
Authentication
Google OAuth↗
Sign in with Google accounts via OpenID Connect, including native Android sign-in with server-side ID-token verification
Microsoft OAuth↗
Azure AD and Office 365 account authentication
Apple OAuth↗
Sign in with Apple for iOS and web users
LinkedIn OAuth↗
Professional identity verification and sign-in
JWT + bcrypt↗
Token-based session management and password hashing
Web Crawling
Playwright↗
Headless Chromium automation for JavaScript-rendered government sites, for browser-testing our own pages, and for rendering the frames of our narrated videos
BeautifulSoup↗
HTML parsing and DOM traversal for content extraction
Trafilatura↗
Main content extraction, stripping navigation and boilerplate
Async HTTP client
Async HTTP client with HTTP/2 support for fast page fetching
Tinyproxy
Authenticated proxy pool across distributed relay servers with weighted load balancing by transfer quota
AI Infrastructure
Claude↗
Anthropic's Claude Haiku for basic image analysis
Voyage AI↗
voyage-3 text embeddings for semantic deduplication of incoming service requests
PyTorch↗
Deep learning framework powering OCR, speech, and embedding models on CUDA
Ollama Embeddings↗
Generates document embeddings via nomic-embed-text for vector search
Ollama LLM↗
Self-hosted inference for text classification, entity extraction, report analysis, and meeting summaries
Sentence Transformers↗
CrossEncoder re-ranking for improved search result relevance
Transformers↗
Hugging Face model runtime and Hub powering OCR and re-ranking, and distributing the speech models
Ultralytics YOLOv8↗
Object detection model for image safety classification
BERTopic↗
Topic modeling that clusters grand-jury findings and recommendations into recurring themes
Document Processing
Surya OCR↗
GPU-accelerated text recognition for scanned government documents
PyMuPDF↗
PDF rendering, text extraction, and page-level manipulation
pdfplumber↗
Fallback text extraction when PyMuPDF output quality is low
pypdfium2↗
PDF page rendering to images for the Surya OCR pipeline
pypdf↗
Page-level structural work: splitting, merging, and extracting from consolidated filings
WeasyPrint↗
Renders print-ready PDFs from the same HTML as the web pages: reference sheets, papers, and reports
Speech Processing
FFmpeg↗
Audio and video decoding, 16 kHz mono audio extraction from meeting recordings, and encoding of our narrated videos
WhisperX↗
Word-level speech-to-text transcription of public meeting audio
pyannote.audio↗
Speaker diarization: segmenting who is speaking and when
SpeechBrain↗
ECAPA-TDNN speaker embeddings for voice attribution across meetings
Kokoro-82M↗
Open-weight text-to-speech model that voices the narration of our explainer videos, run locally on the CPU
Geospatial
Leaflet↗
Interactive maps on entity, service, and county pages, fed simplified GeoJSON scaled to the zoom level
Esri ArcGIS Online↗
Street, terrain, and aerial imagery tiles under every map we render. Tiles © Esri
Shapely↗
Geometry operations for boundary processing: dissolving overlaps, clipping, reprojection checks, and spatial analysis
GeoPandas + pyproj↗
Reading national boundary files from shapefile, GeoPackage, and GeoJSON, joining them to entities, and moving them between coordinate systems
Census TIGER/Line↗
U.S. state, county, municipal, and school-district polygons joined by GEOID, plus the road network behind corridor and ownership queries
USGS PAD-US↗
The national inventory of protected and public lands, carrying owner and manager as separate attributes so a report reaches whoever runs the place
CPAD↗
The California Protected Areas Database from GreenInfo Network, the in-state authority that feeds PAD-US
GADM↗
Administrative boundaries for first- and second-level subdivisions across the countries we cover
Natural Earth↗
Land and water masks used to clip assigned water back out of shoreline boundaries
OpenAddresses↗
Open address points behind typed-address autocomplete, loaded statewide for California
ArcGIS REST↗
State and county open-data feature services, from special-district boundaries to live parcel lookups, with per-county field mapping
OpenStreetMap↗
Boundary gap-fill, private road and land layers, and the data behind Nominatim
Nominatim↗
Forward and reverse geocoding for jurisdiction routing, with country-aware dispatch to regional geocoders
Census Geocoder↗
U.S. address geocoding, with a per-state bounding-box acceptance check on every result
Communications
SMS and RCS gateway
SMS and RCS messaging for alerts, OTP verification, notifications, and message relay
Transactional email provider
Transactional email delivery for user notifications and digests
Email Relay Network
Dedicated mail servers for two-way correspondence with government agencies (FOIA, public records, service requests)
Postfix + Rspamd↗
Mail transfer with spam filtering, DKIM signing, and rate limiting
SPF + DKIM + DMARC↗
Sender authentication with enforcement policies on every UnGovr domain, plus automated monitoring of DMARC aggregate reports for spoofing attempts
ClamAV↗
Inbound attachment scanning for malware detection
Cloudflare Email Routing↗
Inbound addresses on our site domains routed to staff inboxes, with an Email Worker that delivers one address to several people
Payments
Mobile
React Native↗
Cross-platform mobile app for service requests and reporting
Expo↗
Build toolchain with camera, location, and push notification integration, plus over-the-air updates for JavaScript changes between store releases
TypeScript↗
Type-safe JavaScript across the mobile app and LexCode
Edge & CDN
Project Galileo↗
Cloudflare's Project Galileo provides free, robust security to vulnerable Internet properties that are the targets of DDoS and other cyber attacks. As a government transparency platform, UnGovr is proud to be a Galileo-protected site, receiving enterprise-level DDoS mitigation, DNS, SSL, and CDN services at no cost.
Cloudflare DNS & SSL↗
Domain management, DDoS protection, and universal TLS certificates
Cloudflare Zero Trust↗
Secure origin connection without exposing server ports
Cloudflare Workers↗
Edge routing, request rewriting, and error handling at the CDN layer, plus standalone Workers behind the Open Data API, the public status page, and shared LexLint results
Cloudflare R2↗
Zero-egress object storage for the static sites, the Open Data API's files, uploaded service-request media, and off-site database backups
MCP Servers↗
Model Context Protocol endpoints on Workers. One gives AI agents structured access to entity data and grand jury reports, settling per-request payments via the Machine Payments Protocol over HTTP 402. The other lets coding agents lint an app against LexLint's law library
Open Data API→
Entity data and grand jury reports as JSON files served from R2 by a Worker at data.ungovr.org, with API keys held in Workers KV and usage counted in Analytics Engine
Cloudflare Durable Objects↗
The prepaid API wallet's ledger: one object per account, so every top-up and every charge is recorded exactly once, in order
Pagefind↗
Static full-text search, indexed when a site is published and served from the CDN with no search server, on our oversight and observatory sites and the LexLint developer docs
Cloud Infrastructure
Akamai Linode↗
Cloud compute in Los Angeles hosting UnGovr's interactive tier, built so logins, browsing, and service requests stay available independent of the origin server
WireGuard↗
Encrypted network link carrying database replication and internal traffic between the origin server and the cloud tier
Logical Replication↗
PostgreSQL logical replication streaming entity, geography, and routing reference data to the cloud tier in near real time
Hetzner↗
A separate cloud region providing redundancy across UnGovr systems, from off-site database backups to standby mail-relay capacity, kept independent of the primary region
Code Analysis
Opengrep↗
Open-source static analysis engine behind LexCode, our tool that records, file and line, what a repository's code does in LexLint's activity words. Each release is pinned by checksum
Node.js↗
Runtime for the LexCode command line, on a developer's machine and in CI
npm↗
Registry LexCode is published to, so a developer runs it with one command
Vitest↗
Test runner holding LexCode's rules, manifest readers and drift check to their expected results











