AI Meets a 96-Year-Old NGO: Transforming Cross-Border Child Protection with Smarter Case Management

Why This Project?

ISS manages over 70,000 open cases every year, involving vulnerable children and families across borders. This enormous volume of sensitive, unstructured information creates several operational challenges:

• Critical details are often buried in lengthy documents

• Manual data entry consumes hours of caseworker time

• Staff turnover weakens knowledge continuity

• Delays in information retrieval slow down case progress

Reducing repetitive documentation could free caseworkers to engage in meaningful client-facing work. ISS recognized that modernizing its case management tools was essential — and turned to Omdena to co-create an AI-driven solution tailored to nonprofit realities.

Problem Statement

The project aimed to improve how ISS manages and processes case information by reducing administrative workload and improving data usability.

To achieve this, the first step was to understand the specific factors preventing caseworkers from focusing on client-facing work. Through discussions and early analysis, our team identified several key challenges, such as repetitive data entry, lack of data organization, and limited tools for quick case retrieval.

These challenges created a cycle of inefficiency. Valuable time was spent managing files rather than directly supporting families in need. It became clear that better data organization could lead to faster case resolution and a higher quality of service.

By improving data management, we could free up time, simplify workflows, and give caseworkers the ability to focus on what truly matters: protecting and supporting children. To demonstrate this concept, we decided to create a sample database system that would show how structured data could transform daily case handling.

Initial Challenge: Lack of Data

Strict confidentiality meant the team could not access real ISS case files. Only five anonymized files were initially provided — far too few for training AI models.

To work around this, the team curated additional case material from public sources, including:

• ISS partner websites
• Government and child protection reports
• Training manuals
• Public court records (ECHR, US Case Law, LawCite)

This resulted in a curated dataset of 230 cases, enabling meaningful modeling without compromising privacy.

The prototype needed to be intuitive for non-technical social workers, simple to navigate, and transparent in how it generated insights.

Solution Overview

Our team collectively tackled the main problem from various angles, as follows:

Graphic by Bima Putra Pratama

The team developed an end-to-end prototype that combined NLP, machine learning, and user-friendly visualization into a single web-based tool. Its goals:

• Reduce time spent reading and summarizing case files
• Predict levels of risk for faster prioritization
• Extract key abuse patterns and risk factors
• Recommend similar past cases for reference
• Provide a structured, searchable database

The platform delivered immediate value while laying the foundation for ISS’s broader digital transformation.

Collecting the Data

Due to limited cases available at the start of the project, the first task in hand was to collect more children related cases from various sources. We majorly concentrated on child abuse and migration cases.

We gathered the case files and success stories that were publicly available on ISS partner websites, Malawi’s Child Protection Training Manual, Bolt Burdon Kemp, and Act For Kids. We also collected a catalog of child welfare court cases from the European Court of Human Rights (ECHR), WorldCourt, US Case Law, and LawCite. In the end, we had managed to collate a dataset of about 230 cases and were ready to utilize these in our project pipeline.

Structuring the Data

To build a reliable AI model, we adopted a supervised learning approach focused on predicting risk scores for each case. Every case was manually assigned a score between 0 and 1, where higher scores indicated higher urgency or greater risk to the child’s wellbeing. The labeling was done by three collaborators, and the final score for each case represented the average of their independent assessments. This ensured consistency and reduced personal bias in labeling.

The risk scores were based on multiple factors, including the presence of past abuse, access to education, caretaker stability, and willingness to care for the child. To enhance the dataset, we added attributes such as case duration, type of service, and country of origin. These additional fields helped structure the data for easier analysis and later database integration, giving caseworkers a more dynamic way to search and filter information.

Finally, we demarcated the risk scores into three categories, using the following threshold.

By the end of this stage, we had a detailed and balanced dataset capable of supporting both statistical analysis and AI modeling. It served as a solid foundation for building predictive algorithms that could assist caseworkers in prioritizing their efforts efficiently.

Preparing the Text

Next, we moved to data preprocessing which is crucial in any data project pipeline. To generate clean, formatted data, we implemented the following steps:

• Text Cleaning: Since the case texts were pulled from different sources, we had different sets of noises to remove, including special characters, unnecessary numbers, and section titles.
• Lowercasing: We converted the text to lower case to avoid multiple copies of the same words.
• Tokenization: Case text was further converted into tokens of sentences and words to access them individually.
• Stop word Removal: As stop words did not contribute to certain solutions that we worked on, we considered it wise to remove them.
• Lemmatization: For certain tasks like keyword and risk factor extraction, it was necessary to reduce the word to its lemmatized form (eg. “crying” to “cry,” “abused” to “abuse”), so that the words with the same root are not addressed multiple times.

Extracting Features

Once the data was cleaned, the next step was to convert the text into numerical form so that machine learning models could process it. We tested four different embedding methods: TF-IDF, Doc2Vec, Universal Sentence Encoder (USE), and BERT. Each of these techniques captures text meaning differently, allowing us to compare their performance. To make the analysis efficient, we reduced vector sizes to 100 dimensions using Principal Component Analysis (PCA) and grouped similar cases using hierarchical clustering.

The team used a dendrogram to determine the ideal number of clusters and evaluated model quality through Silhouette scores. Among all methods, the Universal Sentence Encoder (USE) achieved the best balance of accuracy and interpretability. This embedding model became the foundation for subsequent analyses, as it effectively represented semantic similarities between child protection cases.

By completing this stage, we created a solid framework for text understanding. The embeddings enabled our models to identify patterns, similarities, and key relationships across hundreds of case documents with far greater accuracy than manual review.

Models & Algorithms

Text Summarization

Multiple pre-trained extractive summarizers were tried, including BART, XLNet, BERT-SUM, and GPT-2 which were made available thanks to the HuggingFace Transformers library. As evaluation metrics such as ROUGE-N and BLEU required a lot more reference summaries than what we had, we opted for relative performance comparison and checked for the quality and noise level of each model’s outcomes. Then, inference speed played a major role in determining the final model for our use case, which was XLNet.

Keyword & Entity Relation Extraction

Keywords were obtained from each case file using RAKE, a keyword extraction algorithm that determines high-importance phrases based on their frequencies in relation to other words in the text.

For entity relations, several techniques using OpenIE and AllenNLP were tried, but they each had their own set of drawbacks, such as producing instances of repetitive information. So we implemented our own custom relation extractor utilizing spaCy, which better-identified subject and object nodes as well as their relationships based on root dependency.

Similarity Clustering

The pairwise similarity was computed between a given case and the rest of the data based on USE embeddings. Among Euclidean distance, Manhattan distance, and cosine similarity, we chose cosine similarity as our distance metric for two reasons.

First, it works well with unnormalized data. Second, it takes into account the orientation (i.e. angle between the embedding vectors) rather than the magnitude of the distance between the vectors. This was favorable for our task as we had cases of various lengths, and needed to avoid missing out on cases with diluted yet similar embeddings.

After getting similarity scores for all cases in our database, we fetched top five cases that had the highest similarity values to the input case.

Risk Score Prediction

A number of regression models were trained using document embeddings as input and manually labeled risk scores as output. Tensorflow’s AutoKeras, Keras, and XGBoost were some of the libraries used. The best performing model — our custom Keras neural network sequential model — was selected based on root mean square error (RMSE).

Comparison of risk prediction model accuracies

Abuse Type & Risk Factor Extraction

To make the system more insightful, we developed specialized tools that automatically identified patterns of abuse and underlying risk factors within case texts. The team defined eight categories of abuse-related verbs, including words like beat, molest, and neglect, to capture the nature of each case. Using spaCy’s pre-trained English model and part-of-speech tagging, verbs from the text were converted into vectors and compared with these categories using cosine similarity. This helped determine which abuse types were most relevant to each case.

The same method was extended to risk factor detection. We introduced a broader set of keywords such as trauma, sick, war, and lack to identify additional risk indicators. Each word in the case text was analyzed against this list, and those with a similarity score above 0.65 were extracted and presented in their original form. This process provided greater transparency into which words most influenced the assigned risk scores.

By combining these features, the system could generate both quantitative and qualitative insights, allowing caseworkers to understand not just the level of risk, but the specific factors contributing to it. This level of detail improved interpretability and trust in the model’s outputs, helping social workers make more informed and empathetic decisions.

Web Application

To put these models altogether in a way that ISS caseworkers could easily understand and use, a simple user interface was developed using Flask, a lightweight Python web application framework. We also created forms via WTForms and graphs via Plotly, and let Bootstrap handle the overall stylization of the UI.

A Javascript code to implement Google Translate API was incorporated into the HTML templates, enabling the translation of any page within the app into 108 languages.

For the database, we used PostgreSQL, a relational database management system (RDBMS), along with SQLAlchemy, an object-relational mapper (ORM) that allows us to communicate with our database in a programmatic way. Our dataset, excluding the five confidential case files initially provided by ISS, was seeded into the database, which was then hosted on Amazon RDS.

The seeded database also includes fields like summary, risk score, and other model outcomes

A public Tableau dashboard to visualize the case files was also added, should caseworkers wish to refer to external resources and gain further insight on case outcomes.

Conclusion

The goal of this project was to support ISS in delivering faster and higher-quality services by integrating AI into its case management process. Within just eight weeks, the Omdena team successfully developed a working prototype that showcased how Natural Language Processing (NLP) could analyze case data, predict risk levels, and simplify decision-making. The tool demonstrated the potential of technology to enhance the efficiency of social workers while keeping the focus on the children and families they serve.

More importantly, this project proved that even long-established NGOs can benefit from adopting AI in meaningful, ethical ways. The collaboration between ISS and Omdena became a stepping stone in ISS’s digital transformation journey, setting the stage for future applications of data science in humanitarian work. To understand how other nonprofits can replicate similar success stories, explore our complete roadmap on how to build and implement AI products in the NGO sector.

Limitations

Like any early-stage AI initiative, this project faced several limitations. The dataset was relatively small and contained only two service categories: child abuse and migration. Privacy restrictions also limited access to real case files, which meant the models were trained on publicly available data. These constraints introduced some level of bias, especially in the manually labeled risk scores that relied on the team’s assumptions.

Despite these challenges, the project delivered a functional proof of concept that achieved its main goal — demonstrating how AI can enhance the case management process without compromising confidentiality. The system is not designed to replace human judgment but to complement it, allowing caseworkers to make better-informed and faster decisions.

Moving Forward: AI and case management

To move this prototype closer to production, the next step is to incorporate official ISS data in a secure environment to refine model accuracy. Introducing privacy-preserving techniques such as differential privacy could enable safe data sharing while maintaining confidentiality. Continuous learning through caseworker feedback can further improve the accuracy of risk scoring and pattern recognition.

Expanding the database fields to include attributes like background information, case outcomes, and service type would provide richer insights and better search capabilities. Once deployed on a scalable platform like AWS, the application could be used by ISS caseworkers worldwide, connecting them through a shared data-driven ecosystem. This progress would empower social workers globally — enabling them to spend less time on paperwork and more time protecting and supporting children.

This article is written by Shrey Grover and Jianna Park.