How Justt Scaled Chargeback Document Extraction Using NVIDIA Nemotron Parse

Challenge

What Justt needed to do

Justt set out to modernize its unstructured data extraction pipeline to deliver high-fidelity results for customers interacting directly with the platform. Because users depend on real-time outputs to manage chargebacks, even small inaccuracies from legacy systems required costly manual corrections and disrupted workflows. Traditional CPU-based approaches struggled to accurately interpret complex document layouts and semantic relationships at scale. As document volumes grew, the system hit strict latency ceilings and rising infrastructure costs, making it clear that a fundamental architectural shift was required to achieve both accuracy and efficiency.

Constraints

• Infrastructure: The existing pipeline was built on CPU-based OCR workflows, primarily using the community library. While effective for simple documents, this approach struggled with complex layouts and dense semantic relationships, limiting both throughput and extraction accuracy. Supporting modern deep learning–based inference required a transition to a GPU-accelerated architecture.

• Performance and latency: Justt’s customers interact directly with the platform and expect immediate results. To support time-sensitive chargeback workflows, the extraction pipeline needed to deliver near-real-time performance while operating within strict end-to-end latency limits.

• Cost efficiency: The system processes large volumes of documents daily, making cost per document a critical factor. Although GPU acceleration offered significant performance benefits, the solution had to maximize utilization and efficiency to meet the economic constraints of a production-scale, enterprise deployment.

System architecture

To meet the performance, accuracy, and cost constraints, Justt designed a GPU-accelerated document processing architecture optimized for high-throughput inference. The system integrates Nemotron Parse, part of the NVIDIA Nemotron™ family, to enhance visual document understanding, enabling robust extraction across complex layouts while preserving semantic relationships between fields.

Nemotron Parse standardizes context-aware field mapping across document types, allowing extracted data to be normalized into high-fidelity outputs without manual post-processing. By running inference on GPUs, the architecture achieves low-latency processing while supporting efficient scaling as document volumes increase.

Click Image to Enlarge

Figure 1. Justt’s intelligent document processing pipeline design includes an OCR engine and AI-powered extraction component for visual and text extraction and analysis.

Architecture overview and data flow

Integration details

Key design decisions:

Justt focused on adapting its intelligent document processing pipeline to support multimodal inputs and outputs along with how it was deployed.

• Pipeline adaptation: Integrating Nemotron Parse required modifying multiple components in the extraction pipeline to support the model’s input and

• Model serving architecture: Due to hardware requirements for GPU-accelerated inference, Nemotron Parse was deployed as a dedicated model-serving service on AWS, decoupled from the core platform application. This separation enabled independent scaling of inference workloads without impacting application-layer performance.
  

• Evaluation and quality measurement: The evaluation framework was updated to reflect the new architecture, introducing revised metrics to measure extraction accuracy and error rates produced by the GPU-accelerated pipeline.

How Nemotron Parse was integrated

• Inference trigger and preprocessing: Nemotron Parse is invoked as part of the document ingestion workflow. Upon customer file upload, documents are converted into page-level image representations and passed to the model for inference.

• Post-processing and field alignment: The model’s outputs are forwarded to an LLM-based processing layer that performs domain-specific field alignment, translating extracted content into Justt’s standardized schema.

What didn’t work initially

Initial installation required dedicated time to set a new environment and adjust the installation to the relevant hardware profile early on. To use Nemotron Parse, the model swap was straightforward and did not require additional optimization or customization.

Results

Measured outcomes

Transitioning from legacy OCR-based extraction to Nemotron Parse resulted in a 25% reduction in extraction error rate across evaluated document sets. The improved accuracy reduced the need for manual corrections during document uploads and increased the reliability of downstream chargeback evidence generation.

Customers such as HEI Hotels & Resorts, a leading hospitality operator, use the platform powered by this architecture to process chargebacks efficiently across multiple properties while improving recovery rates.

Before and after comparisons

• 25% fewer manual corrections: Customers spend significantly less time correcting extracted data during the upload process, enabling faster case preparation and reduced operational overhead.

• Higher-quality evidence generation: By minimizing false positives and human error at the extraction stage, the system produces cleaner, more reliable data for chargeback defense workflows, contributing to improved revenue recovery outcomes.