DigitalSafety

Inspiration

The ultimate inspiration for this project comes from organizations like Thorn and International Justice Mission.

What it does

DigitalSafety is an AI-assisted system designed specifically for low-resource environments dealing with incomplete, highly sensitive risk indicators. Rather than assuming clean data or unlimited staff, the project maps the operational flow from Signal to Outcome. The core operational pipeline include Input Signal: processes only the 6 abstract risk signals: Signal Anomaly Count,Cross Platform Sync Rate,Encrypted Message Frequency,Flagged Keyword Density,Account Age in Months,XAI Feature Weights. Predictive Model (Linear Regression): These risk signals are fed directly into the locally trained Linear Regression Model. The model instantly calculates the Investigator Priority Score (0–100) based on optimized coefficients. Generative Language Model (ollama): The calculated Priority Score and the risk signal valuse are bundled into a secure, structured prompt template. This prompt is sent to the local Ollama API running entirely on the organization's offline hardware. The LLM translates the raw mathematical metrics into plain-text Actionable Insights. outcome: The human investigator makes their final decision (either accepting the system's recommendation or manually overriding it).

How we built it

the project was built using python programming language, keras, pandas, numpy, google-ml-edu libraries were used for the development of the linear regression model. ollama API was used coupled with a prompt to generate an actionable insight and recommendation from the model predicted output. more details on the linear regression model: 1000 synthetic data was used for the model training hyperparameter(50 batch size, 30 epoch 0.1 learning rate), the model achieved loss: 5.8170 - rmse: 2.4118, as it converge. out of these 1000 dataset. 50 random samples were obtained for validation.

Challenges we ran into

getting real world data within the limited time, synthetic data profiles may not capture the constantly changing, real-world tactics used by exploiters.

Accomplishments that we're proud of

We Successfully coupled an objective, Linear Regression model directly with a local Generative Context Layer (Ollama API), proving that quantitative risk scores can be safely translated into plain-text case summaries. We achieved a completely offline, data pipeline that processes highly sensitive metrics without ever exposing a single vulnerable individual's digital footprint to third-party cloud engines.