Where the data comes from
The platform combines several levels of official data. The finest geographical level publicly available in Romania today is the sector (Bucharest is divided into six administrative sectors) - and the only official series geolocated by sector is the crime coefficient published by Poliția Capitalei (the Bucharest Police).
The official layer by sector - the crime coefficient (DGPMB)
Direcția Generală de Poliție a Municipiului București (DGPMB, the Bucharest Police Directorate) publishes each year a specific crime coefficient for each of the 6 sectors, with a risk rating (high / medium / low). It is an official index - not the raw number of offences - and forms the "Official data" layer on the map.
| Sector | 2022 | 2023 | 2024 | 2025 | Trend |
Source: DGPMB - Coeficienți de criminalitate (crime coefficients), b.politiaromana.ro
The national context - verified figures
| Indicator | Value | Year | Source |
| Offences reported nationally | 527,535 (+5.1%) | 2022 | SNOSP 2023-2027 |
| Street crime | +7.3% vs. 2021 | 2022 | SNOSP 2023-2027 |
| Road safety offences | +14.3% vs. 2021 | 2022 | SNOSP 2023-2027 |
| Offences in schools | +83.6% | 2022/2023 | SNOSP 2023-2027 |
| Domestic violence | +56.8% compared with 2016 | 2022 | SNOSP 2023-2027 |
| Protection orders issued | 15,256 (+10%) | 2022 | SNOSP 2023-2027 |
| Offences recorded | 358,960 (+2%) | 2023 | MAI - 2023 results |
| Homicides (falling) | 216 (down from 238) | 2023 | MAI - 2023 results |
| Criminal cases resolved | 502,584 | 2024 | Poliția Română 2024 report |
| Robberies / thefts (IGPR level) | -15% / -5% | 2024 | Poliția Română 2024 report |
Sources:
SNOSP 2023-2027 ·
MAI ·
Poliția Română - statistics ·
data.gov.ro ·
INSSE Tempo Online
SNOSP: Romania's National Public Order and Safety Strategy. MAI: Romania's Ministry of Internal Affairs. Poliția Română / IGPR: the Romanian Police and its General Inspectorate. INSSE: Romania's National Institute of Statistics.
The real incidents layer - from official press releases and the media
Although Romania does not publish an incident-level geolocated dataset, individual events are reported in press releases from Poliția Capitalei (DGPMB), Poliția Română, MAI and ISU (the emergency services inspectorate), as well as in media reports citing the authorities. We collected and manually verified dozens of real cases from Bucharest (2023-2026) and placed them on the map as separate pins. Each pin opens a card with the type of offence, the area, the date and a direct link to the source, so that anyone can check the information.
- Official sources do not publish the exact street (to protect investigations and victims), so precision is at sector / area level - consistent with the anonymisation principle.
- Every incident has a real, verified URL; nothing is made up.
The three types of data on the map
1. Official data by sector (the coloured areas) - the DGPMB coefficient, real. 2. Real incidents (pins) - cases from official press releases and the media, each with a link to the source. 3. Pilot incidents (dots) - a demonstration set showing what the platform will look like at scale, once connected to feeds from Poliția Locală, the local police (Phase 1). For now, the prediction is based on the available layers and is a demonstration model.
Complementary sources integrated
- OpenStreetMap (ODbL) - urban vulnerability factors: police, hospitals, schools, nurseries, transport, nightlife, surveillance cameras. Extracted via Overpass and shown on the map as the "Urban vulnerability" layer. Attribution: © OpenStreetMap contributors.
- INSSE - 2021 census - population by sector, used for density and for rates "per 100,000 residents". At sector level, the census is more accurate than grid-based estimates (e.g. GHS-POP, useful for sub-sector resolution in future stages).
- Eurostat (CC BY 4.0) - offences per 100,000 residents for Bucharest (dataset crim_gen_reg), as a comparable European benchmark. See the Analysis page.
Anonymised by design
No identifiable personal data is displayed or processed for public display. Protection is built into the architecture from the start ("privacy by design", Art. 25 GDPR), not added afterwards:
- Anchoring to an area (snap-to-street). An event never appears at its exact address; it is anchored to the nearest area or landmark - the model validated by the UK data protection authority for police.uk.
- Minimum aggregation threshold. An area is shown only above a minimum number of events. Below that threshold nothing is shown, so that re-identification is impossible.
- No data about people. We do not use arrest data, names, ages or any other individual attributes. The analysis looks at places and patterns, not people.
- Generalisation to block / neighbourhood level of any location information, before publication.
The role of artificial intelligence
A large language model (LLM) works "behind the scenes" to turn heterogeneous reports into clean, comparable data, and to estimate concentrations of risk by area:
- Cleaning and standardisation - bringing data from different formats and classifications into a single taxonomy of categories.
- Classification of events into the platform's categories.
- Identifying patterns and estimating risk by area - where events are likely to concentrate, so that city halls can direct resources preventively. The estimate is by place, never by person.
Processing uses Mistral AI, European artificial intelligence infrastructure, to meet GDPR requirements on data protection and transfers, and the data stays within the EU.
The legal compliance framework
| Framework | What it means for the platform |
| GDPR (Reg. 2016/679) + Legea 190/2018 (Law 190/2018) | The data displayed is aggregated and anonymised; no identifiable personal data. |
| Legea 363/2018 (Law 363/2018, Directive 2016/680) | Governs the processing of data by law enforcement authorities - defines what can be shared. |
| EU AI Act | Prohibits predictive policing based on individual profiling. Permits aggregated, place-based analysis - our model. |
| Legea 179/2022 (Law 179/2022, open data) | Provides the legal basis for opening up and reusing public sector data. |
The compliance conclusion: the safe model is "place-based", aggregated and anonymised - never "person-based". That is the difference between a public transparency platform and a surveillance tool.