Selection of Past Projects


Here you can see a selection of projects that we've developed for our clients.



Automated Invoice Data Extraction





Tags

Accounting & Finance,
Entity Extraction,
AI Document Processing

Category

Natural Language Processing

Client

Decimo GmbH (rechnung.de)


img


Project Description

For our client, decimo GmbH/rechnung.de, we’ve built an AI system that automates their invoice processing tasks, to make their back-office more efficient. Our system is able to detect over 35 categories in invoices, with super-human performance. Whether it’s the name, street and zip code of the invoice issuer or of the recipient, the tax amount, the invoice-item description or the invoice-item amount, our System is able to detect all of the important categories that are required for accounting.

Since our system comes in the form of a developer friendly rest-api, the results can be fed right away into every database- and accounting-system like Datev, SAP etc. The system takes an invoice as input (whether as an image or PDF) and returns the detected categories in JSON format. Optionally, it can also return a visualization of the categories on the input-invoice.

We’ve only worked with state-of-the-art Artificial Intelligence, which enables us to process any invoice right away, without any specific rule or template setup. Since manual document processing is a major cost driver in organizations, the usage of our system results in a massive cost reduction for our client.

  • Superhuman performance

  • 6× speedup of data entry compared to a manual process.

  • Up to 90% cost saving

  • No specific rule or template setup required. We can work with any invoice layout right away

  • Seamless integration into the current workflow and processes of our client
Looking for Professional AI Experts?


Food Detection

Automating the Payment Process in Canteens





Tags

Object Detection,
Deep Learning,
Consulting,
Real Time Detection

Category

Computer Vision

Project Description

This project is the startup idea of three serial entrepreneurs who want to automate the payment process in canteens. The product can be described as follows: The guest moves his tray under the camera. Our system scans the items on the tray and automatically recognizes and quantifies the products in real time. The names of the products, along with their prices, are then displayed on our dashboard. This enables the system to calculate the total price for the customer, who can then start the checkout process and pay for his products.

Challenge: The system should be able to learn the detection of new products only with a few images (e.g. "continous-learning"). Usually, at least a few hundred images are required to train a neural network to detect a new item and training AI systems with only a few images is still an open research field. Because of that, not many engineers are able to build sufficient systems, which is the reason why this part of the project was the biggest challenge.
This specific feature is needed because a canteen offers new dishes every day. Because of that, there is only a limited amount of time and limited data available to train the object detection system on these new dishes.

Through our experience in building AI systems, we where able to solve the "continous-learning" challenge and deliver an accurate and robust object detection system, together with a responsive web-dashboard to visualize the results.

  • Real time detection

  • Dashboard to show detections, prices etc.

  • Learns to detect new products with only a few images

  • Accurate & robust detections

  • Easy to use (you only need a webcam and a laptop)

Looking for Professional AI Experts?


Lymph Node Cancer Detection





Tags

Image Classification,
Deep Learning,
Healthcare,
Convolutional Neural Networks

Category

Computer Vision

Project Description

In this project, we created a deep learning system to identify metastatic lymph node tissues in digital pathology scans. We also built a dashboard to show the diagnosis of our system, the probability of the diagnosis, and visualizations of why the model made a certain decision. It shows the human doctor what parts of the input image lead to the diagnosis, but also what attributes of the image speak against it. Additionally, it provides the doctor with the top 12 feature maps, along with their relative impact on the diagnosis. With this information in place, the doctor can understand what makes the model think that this person has cancer or not and he can decide with human judgement if he agrees with the diagnosis. It provides him with a fact-based diagnosis to prevent subjective reasoning and to speed up the diagnosis.

We can proudly claim that our system is able to detect lymph node cancer with an accuracy of over 97%, whereas experienced human pathologists can only achieve a performance of 88%. Therefore, this project is a perfect example of the power & the capabilities of deep learning and AI.

  • Outperforms human doctors

  • Enables a fact-based diagnosis

  • Visualize what leads to a diagnosis

  • Shows the probability of a diagnosis

Looking for Professional AI Experts?


Pneumonia Detection





Tags

Image Classification,
Deep Learning,
Healthcare,
Convolutional Neural Networks

Category

Computer Vision

Project Description

Here we had to build an AI system to automatically detect and locate lung opacities on chest radiographs (e.g. pneumonia). Pneunomia is one of the top 10 causes of death in the United States and diagnosing the disease is a complex undertaking that previously only highly trained specialists could do. They had to take a patients clinical history, vital signs and laboratory exams into account, which made the whole process very time consuming.
Our system is there to assist the doctor in diagnosing pneunomia. It is able to make accurate & robust diagnoses, only given the chest radiographs. It provides him with a fact-based diagnosis that prevents subjective reasoning and significantly speeds up the time it takes to make a diagnosis.

For this project, we also build a dashboard that has the same capabilities as the dashboard from the "Lymph Node Cancer Detection" project (see section above), but it also displays meta-information about the patient. With it, the doctor can understand why our system made a certain diagnosis and decide with human judgement if he agrees with it.

  • Enables a fact-based diagnosis

  • Shows what leads to a diagnosis

  • Shows the probability of a diagnosis