Oak-Tree Technologies

Background AdPartner is a European advertising network that uses content classification to match ads to interested buyers based on categorical and contextual signals. Oak-Tree created a classification model and pipeline to categorize content based on theme and subject matter. The Challenge The model needed to be able to cope with extremely high load (over 5000 queries per second) and be able to scale on demand. Each piece of content moving through the pipeline needed to be individually parsed, translated, categorized, and delivered back into the client’s data systems to become part of an ever-growing database. The Solution Using a data streaming architecture based on a serverless approach under Open-FAAS, we created a system for retrieving URLs for translation, parsing, and cateogirzation. All steps in the pipeline were implemented as serverless functions. The pieces of the architecture communicated with one another using Kafka.The database was implemented using Citus (a distributed PostgreSQL technology) with a custom API and UI implemented using Django. Batch processes which started and monitored the pipeline were implemented using Apache Airflow.

Background Every industry wants computer vision. Image classification, video inference, or 3D reconstruction are examples of computer vision that can automate day-to-day tasks with precision. One of the more popular requests we get from clients is to provide programmatic object detection from a given video/image. The Challenge Most object detection use cases are for smaller, portable devices for live video feeds. As object detection requires deep learning, the processing power in these devices must have a strong enough computing ability to detect objects in real-time. The Solution Through deep learning TensorFlow libraries that specialize in object detection, we write models that successfully identify objects with given videos. Our solution involves importing our software into a mini GPU computer such as the NVIDIA Nano Jetson. These devices are capable of immediately collecting real-time video feed and applying accurate object detection.

Background While English is one of the more difficult languages to learn, it has the highest demand worldwide. By harnessing the power of AI, teaching English can be scalable and automated. The Challenge Building virtual assistants that can replace the job of a human teacher takes in heavy requirements. Overall, this application must process various non-native English speaker’s voices and writing and apply suggestions to perfect the grammar, pronunciation, and punctuation. The processors must handle dozens of different languages to teach students globally and provide a web UI for students to visualize their learning environment. The Solution Through utilizing Natural Language Processing (NLP), the linguistics field of artificial intelligence, we can process and analyze linguistic data, whether written or spoken. We produced a suite of AI tools centered around interactivity with the user from NLP libraries and data operations. The application can hear fragmented English sentences and correct them to a natural-sounding English with labeled errors so that students can know what mistakes they made. The Result The final result successfully deployed an ensemble of models that contain TensorFlow and PyTorch models surrounded by API systems. The application is displayed as a web interface where students can see their English results and improve from all levels.

Sonador is an open source cloud platform for AI-powered applications which use medical imaging data. It includes SDKs for storing, visualization, and quantifying data to create real-time, secure, and scalable solutions to address clinical challenges.