Nexius AI: Lessons from a Financial Document Processing System
What I learned while working across OCR, parsing, validation, background processing, reporting, and delivery in Nexius AI.
Muhammad Gilang Ramadhan
Software Engineer
Product Scope
At Quantum Teknologi Nusantara (September 2025 – May 2026), I worked on Nexius AI, a financial document processing product that uses OCR and AI-assisted workflows. The platform converts uploaded documents into structured outputs that users can review and download.
The product spans customer-facing upload flows, partner and affiliate portals, and internal admin dashboards, so engineering decisions need to support both product usability and operational visibility.
Backend Architecture
I standardized FastAPI service architecture using Domain-Driven Design so services remain maintainable as AI integrations and document workflows grow.
Long-running file-processing flows were moved from monolithic processing into RabbitMQ-based distributed workers and Kubernetes worker pods to improve scalability under high upload volume.
- REST APIs for uploads, queue visibility, report/month processing state, and background jobs.
- Server-Sent Events for real-time upload and processing progress.
- Worker heartbeat, queue diagnostics, and progress snapshots for operational debugging.
AI Quality and Observability
The AI workflow includes transaction categorization, ambiguous transaction clustering, Chart of Accounts mapping, and metric-based validation.
I also improved observability with OpenTelemetry, SigNoz, structured logs, and queue diagnostics so production issues are easier to trace.
written by
Muhammad Gilang Ramadhan
Software engineer building distributed backend systems and applied AI products, with a background in competitive programming.