Computer Vision & Distributed Document ProcessingThesis (R & D)
PROJ // 01
DigiDoc:
A Privacy-Preserving Distributed Document Digitization Framework for Automated Form Transcription (SaaS)A privacy-focused document processing framework that implements the following features:
- Designed and developed a secure document digitization platform with role-based access control (RBAC), JWT authentication, and OAuth2-based authorization.
- Developed a document processing pipeline involving template alignment, field annotation, coordinate-based segmentation, and sensitivity scoring.
- Developed the NDT (Non-reproducible Document Digitization and Transcription) algorithm with O(n) time complexity, implementing sensitivity-aware segment distribution, workload balancing, and anti-collusion constraints through an n-th independent verifier distribution mechanism that assigns the same segment group to multiple independent verifiers while keeping their assignments and transcription results isolated, preventing individual verifiers from accessing or reconstructing the complete original document.
- Integrated MiniLM v2 embeddings and semantic similarity scoring to automatically group related fields, reduce manual annotation workload, and improve document understanding.
- Integrated OCR-assisted text extraction with human-reviewed transcription, evaluating performance using Character Error Rate (CER) and Word Error Rate (WER).
- Implemented a robust backend using FastAPI with PostgreSQL and SQLAlchemy ORM, and utilized Redis and Celery for asynchronous background processing and caching.
- Containerized the application using Docker with Docker Compose for streamlined development, testing, and deployment workflows.
PythonFastAPIPostgreSQLAnnotorious.jsSQLAlchemyRedisCeleryOCRMiniLM v2Computer VisionDockerPrivacy-Preserving Data Processing
