SafeShelter

The Problem

In Singapore, over 500 individuals are experiencing homelessness, often referred to as rough sleepers. To provide temporary refuge, the government and community partners operate 22 Safe Sound Sleeping Places (S3Ps) and 6 transitional shelters, which serve as intermediate housing solutions while applicants await long-term public rental flats.

Despite the availability of support, homeless individuals face long waiting periods, which can severely discourage them from seeking help. For instance:

• The S3P application process typically takes over a week.

• Transitional shelter applications, managed by MSF and its partners, often take 4 to 6 weeks to process.

• Access to subsidised public rental flats may take up to a year.

These delays are not just due to high demand, but also stem from manual bottlenecks in the case evaluation process. Social workers are required to review various case documents (including medical reports, social assessments, financial statements, and handwritten letters) often in non-standard formats like scanned PDFs, photos, or poorly formatted documents.

Although social workers are committed and compassionate, the manual extraction and cross-checking of information against complex eligibility rubrics is laborious and time-consuming. These administrative burdens reduce the time available for direct client support and contribute to the overall delay in placing rough sleepers into shelters.

Our Solution

To address the bottleneck faced by social workers in transitional shelters who must manually review case documents for rough sleepers, we propose SafeShelter — an AI-powered assistant that streamlines the document assessment process.

SafeShelter combines Optical Character Recognition (OCR) and Large Language Models (LLMs) to automate and accelerate the extraction, analysis, and summarization of relevant case information.

Social workers can upload multiple types of documents (including PDFs, images (JPG, PNG), and .docx files) and optionally tag them by document type (e.g., medical report, financial statement, social service letter). The system then performs three key steps:

1. Text Extraction
   OCR is used to extract legible text even from scanned documents and photos, ensuring compatibility with non-digital or handwritten submissions.

2. AI-Powered Summarisation
   The extracted text is processed using an LLM that generates a summary tailored to the shelter’s specific assessment criteria. This includes indicators of financial eligibility (e.g. income, assets), medical or social needs, and urgency of placement.

3. Interactive Editing via Natural Language
   The output is presented in an editable, human-readable format. Social workers can review and refine the summary through a natural language chat interface, requiring no technical training or prompts.

By removing the need for manual review of each document, SafeShelter reduces administrative burden, increases consistency in evaluations, and most importantly, shortens the time it takes for rough sleepers to be placed into transitional shelters.

Meet the team

Team SafeShelter

Team members:

• Xin Suen - PM

• Rachel - Engineer 1

• Danial - Engineer 2

• LL - Engineer 3