All Torn Up: Korean AI Startup Tears Document Digitization Times to Shreds

Using NVIDIA DGX-1 and CUDA, AkuoDigital helps customers cuts paper storage costs by 30 percent.
by Lauren Finkle

The concept of the paperless office arrived in the 1970s. Nearly five decades later, office workers are still using 10,000 sheets of copy paper every year on average, according to the University of Illinois.

AkuoDigital, a Korea-based startup, is using AI to help us pare back on paper products. The company’s technology eases the scanning and digitizing of documents, making the content more accessible and slashing paper storage costs.

The Kindest Paper Cut

Governments and industries like finance, healthcare and insurance revolve around paper-driven practices. Physical storage is expensive and requires regular maintenance, so many businesses turn to scanning as an alternative. However, traditional scanning is often expensive and time-consuming.

The companies that have been willing to spend resources on this kind of digitization often get frustrated with the inaccessibility of their scanned documents. That’s because the paper records are frequently turned into images rather than PDFs, meaning that they’re not easily searchable.

AkuoDigital uses its AI system to rapidly scan and search these documents. However, they had to train the system to overcome a troublesome issue: handwriting. Whereas typed papers were uniform, each handwritten document is unique and could result in poor assessments of what’s on the page.

Moreover, most of AkuoDigital’s customers are based in Korea, which presented an additional problem. In AI circles, Korean is considered a “low-resource language” — one that has little data available on which to train deep learning neural networks. To accelerate their training times, AkuoDigital turned to NVIDIA technology.

Using NVIDIA DGX-1 systems and CUDA software, AkuoDigital’s systems have made their deep learning training 11x faster and sped up data transfer times by 15x.

Digitized with Deep Learning

AkuoDigital first addressed the problem of scanning: it takes a business’s documents and scans them in-house, allowing for cheap and efficient mass processing. The company’s AI-based optical character recognition system makes something better than a regular image — it creates an interactive, searchable file.

Companies with strict regulations regarding document transfer can use the AkuoDigital Transformation Service, called TDTS. It can be operated on blocked company intranets and securely stores and retrieves processed documents.

The company also developed a deep learning framework for handwriting recognition that quickly extracts characters. This includes written Korean, which has characters of widely variable lengths, making it particularly challenging.

Thanks to NVIDIA tech, AkuoDigital is making 600,000 pieces of paper searchable daily and processing 120TB of data per month on average. And they’ve been reducing their clients’ document storage costs by an average of 30 percent.

One client, Korea Marine Transport Co., Ltd., deals with 13 million pieces of paper annually. By partnering with AkuoDigital, they’ve cut their scanning time from 1.5 years to 20 days.

Going forward, AkuoDigital CEO Yongseop Kim says his company is working on extending the intelligence of their technology until each document containing handwritten data can consistently be reduced to texts, figures and tables that can be digitized quickly and meaningfully.