At Rise of AI 2019 in Berlin, Aleksandar Stojanovic, Executive Chairman and co-founder of AVA, explains how his company acts like a weather forecasting service for physical risk: crime, terrorism, fire, natural disasters, civil unrest and healthcare emergencies such as epidemics. He stresses that AVA does not do facial recognition or track individuals but analyzes behaviors and locations, serving clients such as the London Metropolitan Police, public transport operators and logistics companies. He describes AVA's Risk Score as combining past, present and predictive security data for a given place, built from internal partner data, sensors and cameras, social media monitoring, and even legally accessible US police radio scanners transcribed with speech recognition. He gives concrete examples: detecting the 2017 London Bridge attack four minutes before official procedures, predicting crime in Los Angeles from ten years of hourly data with an error rate just above 5 percent, and using trust scoring to filter fake or panic-driven social media reports. He also describes ambient audio baselines, gunshot triangulation by sound, and computer-vision aggression detection as further data sources, framing AVA as a system that contextualizes overwhelming data for human operators rather than replacing them.
Key takeaways
- AVA analyzes behaviors and locations rather than identifying individuals, distinguishing it explicitly from facial recognition systems.
- AVA's system reportedly detected the London Bridge terror attack about four minutes before traditional police procedures did.
- Trained on about ten years of hourly Los Angeles crime data, AVA's predictive models run with an error rate just above 5 percent.
- AVA deliberately excludes personal characteristics like gender or religion from its risk scoring for ethical reasons, and follows GDPR principles by design.
- Signals such as hashtag use, account age and gunshot-triangulating microphones help AVA separate genuine eyewitness reports from panic or fake news.
Detailed notes
A weather forecast for crime and risk
Stojanovic introduces AVA, based in Berlin with an R&D office in Novi Sad, as a system that works like a weather forecast, but for crime, terrorism, healthcare-related risk such as epidemics, civil unrest and political risk rather than wind and rain. He emphasizes that, unlike facial recognition systems, AVA does not look at individuals but at behaviors, locations and events, aiming to inform people who need to know that something is happening without identifying who is involved. Clients include the London Metropolitan Police, public transport operators managing day-to-day operations centers, and logistics companies monitoring global supply chains.
Detecting the London Bridge attack early
He recounts that during pilot projects with several London boroughs and city hall, AVA's system detected the London Bridge terror attack about four minutes before it was picked up through traditional police procedures. This result led to an ongoing working relationship with London police and city hall focused on the security of the city as a whole rather than on tracking individuals.
The AVA Risk Score: past, present and predictive
Stojanovic explains that the AVA Risk Score looks at a location's historic security, its present situation such as an active incident, and a predictive view of what could happen next, combining all three the way meteorologists combine historical and current data. The system is context-sensitive to circumstances such as parking an old car versus a new one, but Stojanovic says AVA deliberately avoids individualizing scores by personal characteristics such as gender or religion for ethical reasons, and the system is designed around GDPR principles to avoid privacy problems for customers.
Data sources: from crime statistics to police scanners
AVA draws on partners' internal data, such as decades of police incident history, alongside data from insurers and mobile network operators, sensors and cameras that detect behavior rather than identity, and manually reported incidents. Stojanovic describes training algorithms on around ten years of hourly Los Angeles Police Department crime data to predict incidents such as stabbings for the following day, with an error rate a little above 5 percent. He also cites a London School of Economics study linking air pollution and crime as an example of using correlations even without a full causal explanation. For digital media monitoring, he explains that fire departments now actively track social media because eyewitnesses often photograph and post before calling emergency services, illustrated by a false-alarm incident at Munich's Olympia shopping centre where social media amplified panic. To filter unreliable reports, AVA runs trust scoring on sources, for example noting that real eyewitnesses in danger rarely take time to add hashtags, and checking account age and connections to spot bots. Stojanovic also describes listening to publicly broadcast US police and fire radio frequencies and transcribing the structured codes with speech recognition to build a real-time incident map, noting this is illegal and can lead to jail in Germany. Further sources include mobile network cell load changes as an anonymous, privacy-preserving movement anomaly signal, ambient audio baselines of places like train stations to detect unusual sounds such as gunshots, gunshot-triangulating microphones tested in Barcelona that can identify the type of weapon fired, and computer-vision cameras that detect aggressive behavior such as fighting in real time rather than only providing evidence after the fact.
From repeat burglary patterns to everyday applications
Stojanovic describes a well-known criminological pattern in which burglars return to a home after the owner replaces stolen electronics, and says AVA supports police and operators with interfaces that combine such patterns with additional data. AVA serves situation centers, building owners and insurers or underwriters covering large events, and field staff who otherwise arrive at an incident with no context; it also offers an SDK so other companies can build their own safety products. He closes with everyday applications: cars that avoid routing through unsafe areas, financial markets reacting to incidents near facilities such as Foxconn, and travel or hotel booking that accounts for neighborhood safety.
Quotes
Our system doesn't look at humans or individuals. Our system looks at behaviors in general, it looks at locations and things that are happening.
Aleksandar Stojanovic
We have been able to see that happening four minutes before the traditional procedures and methodologies that the police has been deploying.
Aleksandar Stojanovic
In Germany it's not allowed. If you do that in Germany, you go to jail.
Aleksandar Stojanovic
Questions answered in this talk
What is the AVA Risk Score?
According to Aleksandar Stojanovic at Rise of AI 2019, the AVA Risk Score combines a location's past (historic) security data, its present situation such as an active incident, and a predictive assessment of future risk, similar to how meteorologists combine historical and current weather data.
How does AVA assess security risk without facial recognition?
According to Aleksandar Stojanovic at Rise of AI 2019, AVA does not identify individuals; instead it analyzes behaviors, locations and events from data sources such as police records, sensors, cameras, social media and audio, and explicitly avoids scoring based on personal characteristics like gender or religion.
How can social media help detect emergencies faster?
According to Aleksandar Stojanovic at Rise of AI 2019, eyewitnesses often photograph and post about incidents like fires before calling emergency services, so AVA monitors platforms such as Twitter, Facebook and YouTube and applies trust scoring, for example noting that genuine eyewitnesses in danger rarely add hashtags, to separate reliable reports from panic or fake content.