Normlogic AI reviews documents against the DORA regulation. I worked across the Python preprocessing service, the Java backend, and the review UI.
What I worked on
- Built the Python document-preprocessing service (Docling parsing and OCR) that feeds the RAG-based compliance workflow: Markdown chunking, chunk tagging, embedding generation, vector database integration, retrieval reranking, chunk expiry, and knowledge update and rewrite mechanisms.
- Designed the backend data models and APIs (Spring Boot, Spring Data JPA, PostgreSQL) for workflow orchestration, including workflow, run, and status entities for execution tracking and lifecycle management.
- Implemented event-driven communication with Apache Kafka and real-time progress updates over STOMP and WebSockets, and contributed to event-processing reliability and Dead Letter Queue handling.
- Implemented concurrency-controlled processing of large PDF collections with Python asyncio and semaphores, powering a multi-document AI review workflow with document-specific answers.
- Set up Hibernate Envers for audit and revision tracking of the answering plan, and integrated services with OpenFeign clients.
- Implemented fine-grained authorization with OpenFGA, and a common interface behind which multiple AI providers plug in.
- Built the tabular document-review and compliance workflow UI with React, React Hook Form, TanStack Table, and Redux Toolkit.