Data & ML Methodology

Our system aggregates data from verified human rights sources, primarily the OVD-Info API and Memorial Human Rights Center. Automated scripts continuously synchronize our database to ensure real-time accuracy regarding arrests, sentencing, and prisoner locations.

We employ Natural Language Processing (NLP) to parse unstructured case summaries. This involves identifying legal actors (judges, investigators), categorizing criminal articles (e.g., '207.3 Fake News'), and extracting surveillance technology vendors involved in the arrest.

Location data is standardized and converted to coordinates via the OpenStreetMap Nominatim API for geospatial analysis. Concurrently, case details are translated from Russian to English using Google Cloud Translation services to ensure international accessibility.

We utilize XGBoost classifiers trained on historical data to assign risk probabilities to new cases. The model evaluates factors such as criminal articles, age, gender, and location to predict the likelihood of 'Urgency' (Immediate Action Required) and the 'Risk of Torture' while in custody.

Our Python microservice runs Prophet time-series models to forecast arrest trends up to 90 days into the future. Additionally, we build network graphs linking cases by similarity (shared charges, location, and tactics) to detect coordinated repression campaigns and communities.

Using LLMs (Large Language Models), we provide generative tools for legal professionals. This includes an Affidavit Generator that synthesizes prisoner data with country condition reports to draft support documents for asylum cases, and an automated system for generating 'Statement of Complicity' dossiers.