Citizen
Empowering Citizen with AI for Smarter Emergency Response
QUICK LINKS
Citizen.com
Citizen is the #1 safety app, celebrated for delivering real-time alerts on local emergencies. With over 15 million downloads and coverage in more than 60 cities, Citizen ensures critical information reaches users when they need it most. Launched in 2016, it has notified people to evacuate burning buildings, deterred school buses from nearby terrorist attacks and led to the rescue of kidnapped children and missing people.
They came to us in 2021 to refine their internal data tools, beginning with frontend development and expanding into a long-term partnership. Fast forward to 2025, and we’re still collaborating on new AI and ML initiatives.
The Vision
Citizen's mission is to Protect the World, something they are focused on achieving by enhancing user safety through real-time emergency alerts. Given the potential safety impact of these alerts, they must be as real-time and accurate as possible.
Citizen gathers the data for these alerts through a network of different sources — which are processed by Citizen's 24/7 analyst team — improving data quality for actionable insights, streamlining emergency responses and equipping analysts with the tools they need to manage and interpret data effectively.
The Challenge
Like any data-driven platform, Citizen faced the challenge of filtering out irrelevant information and correlating relevant information. Without technical refinement, analysts struggled with a large volume of police, fire and medical emergency response information. The high amount of background noise and difficulties in mapping parallel events unfolding across the city lead to inefficiencies and increased manual workloads.
Our mission was to cut through the clutter, improve data preparation and enhance model accuracy to ensure Citizen’s analyst team operated smoothly and effectively.
Technologies
Text Filtering/Processing
OpenAI GPT models (GPT-4.0, 4.0-mini, 4-turbo, 3.5-turbo, Babbage-002), fine-tuned using Google Colab and OpenAI API
Speech-to-Text (ASR)
Whisper models (large-v3, large-v2), fine-tuned on Google Vertex AI Workbench (GPU-accelerated)
Data Management
Google BigQuery for training/testing data
Other
ArcGIS (address to coordinates), Google Cloud tools, standard data science stack
70%
Reduction in manual data processing15+
Markets now powered by scalable ML models80-90%
Reduction in emergency radio noise
What we did
We developed an AI-driven pipeline that leverages OpenAI’s Whisper, a powerful speech-to-text model, combined with advanced classification models to transcribe, filter and extract vital information from emergency radio clips. Through fine-tuning OpenAI's LLM models, we have significantly enhanced clip-filtering accuracy and update detection — leveraging multiple models from GPT-3.5 to GPT-4.0. Our team optimized speech-to-text capabilities to handle low-quality audio from emergency transmissions, ensuring accurate data processing.
On the business side, our solutions reduced manual toil, accelerated response times and facilitated Citizen's seamless expansion into new markets. We designed scalable models capable of processing 40% of filtered data, allowing Citizen to lower operational costs while enhancing user engagement with faster, real-time updates.
Big Takeaways
REDUCED NOISE, ENHANCED ACCURACY
Citizen’s efficiency surged with an 80-98% cut in emergency radio noise, keeping analysts laser-focused on what matters most. With OpenAI’s Whisper tech and fine-tuned LLMs, data flows faster and cleaner, slashing manual work and speeding up emergency alerts — all while driving down costs.
Product accuracy hit human-level standards, tackling poor-quality audio without a hitch. City-specific ML models brought data precision up a notch, filtering duplicates, decoding radio speak and extracting critical details for dependable, real-time alerts.
COST SAVINGS & BREAKTHROUGHS
Citizen’s measure of market expansion efficiency saw nearly a 70% reduction in just the first six months of 2024. Powered by our ML automation, this breakthrough allowed Citizen to expand faster without increasing headcount, hitting operational targets at record speed.
READY TO SCALE
Expansion into new markets became a snap — no heavy lifting, no extra hires required. With fine-tuned LLMs managing 40% of filtered data, we automated key analyst tasks and simplified model training across new cities. This scalability helped Citizen grow quickly, meeting demand while keeping emergency response agile and efficient across regions, while also unlocking the human analyst team to focus more heavily on the editorial quality that Citizen is known for.
Measurable Impact
Our AI solutions cut costs, boosted efficiency, scaled seamlessly and delivered real-time emergency alerts.