AI-powered evidence intelligence platform.

During disasters, protests, humanitarian crises, and human rights incidents, valuable evidence is often captured by ordinary citizens through smartphones. However, this evidence usually exists as scattered images, scanned documents, PDFs, or multilingual text, making it difficult for journalists, NGOs, and humanitarian organizations to process quickly. Manually reviewing hundreds of submissions, extracting important facts, translating local languages, and preparing structured reports can take hours or even days. This delay reduces the effectiveness of emergency response and media coverage. Many existing AI tools are designed for general conversations rather than evidence analysis, while enterprise solutions are often expensive, cloud-dependent, and unsuitable for organizations with limited technical resources. Additionally, evidence submitted in regional languages such as Bengali creates another barrier for international reporting and collaboration. There is a need for a lightweight, AI-powered platform that can automatically organize raw citizen evidence into structured, factual, and exportable reports while preserving the original evidence for future reference. Such a system can significantly reduce manual effort and enable faster, more consistent documentation of critical incidents.

It is an AI-powered evidence intelligence platform that transforms raw citizen-submitted evidence into structured, journalist-ready incident reports within minutes. Users can upload photographs, scanned documents, or text-based PDFs through a simple interface. The system automatically determines the appropriate processing pipeline based on the uploaded file type. For images, OCR extracts visible text while a vision model analyzes the visual content. For PDF documents, embedded text is extracted directly, with OCR applied to scanned pages when necessary. The extracted information is then passed to a large language model, which translates Bengali content into English when required and generates a structured analysis. The AI identifies the incident category, extracts locations and dates when available, summarizes the event, highlights key factual points, and produces a concise headline. Finally, a formatted media brief is generated as both structured data and a downloadable PDF using ReportLab. All uploaded evidence, extracted text, AI-generated analysis, and reports are stored locally in a SQLite database, ensuring a lightweight deployment without requiring complex database infrastructure. Duplicate uploads are detected using file hashing to prevent unnecessary reprocessing and reduce computational cost. The platform is designed with a modular architecture, allowing OCR, vision models, language models, reporting, and storage components to evolve independently. This makes the system easy to maintain and extend with additional capabilities such as audio transcription, multilingual support, or cloud storage in future versions. By automating evidence extraction, translation, analysis, and report generation, WitnessBridge AI enables journalists, NGOs, and humanitarian organizations to transform unstructured citizen evidence into consistent, structured, and actionable reports significantly faster than traditional manual workflows. This reduces reporting time, improves documentation quality, and helps decision-makers respond more effectively during critical situations.

Programming Language: • Python Frontend: • Streamlit Database: • SQLite • SQLAlchemy ORM AI & Machine Learning: • OpenRouter API • GPT-OSS-20B (LLM) • Vision-capable OpenRouter model for image understanding • EasyOCR (English + Bengali OCR) Document Processing: • PyMuPDF (fitz) for PDF text extraction • Pillow (PIL) for image processing Report Generation: • ReportLab Utilities: • python-dotenv • Requests • NumPy • hashlib • pathlib • shutil • JSON Development: • Git • GitHub • Puku Editor

• GPT-OSS-20B via OpenRouter API for incident analysis, translation, structured JSON generation, and report creation. • Vision-capable OpenRouter model for image understanding. • EasyOCR for extracting English and Bengali text from images. • AI coding assistant (Puku CLI) was used to accelerate development, generate boilerplate code, and assist with implementation. The project architecture, integration, testing, debugging, and customization were completed by the participant.