ISR & Predictive Analytics

We bring extensive experience developing RL agents for mission management across domains. With our established software framework, we efficiently develop mission scenarios, model targets and sensors, and train agents using a variety of RL algorithms. As a result, we’ve created agents that offer unique benefits for mission management, including the ability to:

• Dynamically react to real-time sensor data.
• Support deployment for on-board applications.
• Learn complex strategies to find hard targets or overcome adversaries.
• Balance real-time decision making with strategic high-level objectives.
• Easily enhance agents by modifying the training environment.
• Leverage self-supervised learning, which does not require labelled data.

Transforming Streaming Spatiotemporal Data into Knowledge and Action in Real-Time

AlphaK is an AI-powered spatiotemporal stream processor inspired by the sophisticated OpenAI Five and Deepmind AlphaStar architectures, both of which demonstrated superhuman performance on complex video games. AlphaK transforms streaming data (outputs from Automatic Target Recognition [ATR]) into detections of anomalous activity and complex behaviors. The model leverages advanced AI techniques, like scatter connections, to efficiently combine spatial and temporal data and deliver powerful situational awareness information.

With a government-generated evaluation dataset, AlphaK demonstrated greater than 95% accuracy for classifying multiple challenging behaviors while delivering extremely low latency. The AlphaK architecture has been deployed on AWS GovCloud with containerized microservices and showed autoscaling to quickly meet mission demands. AlphaK has also been embedded in a Deep RL Framework to transform streaming data into sensor tasking recommendations required to accomplish high-level objectives.

AlphaK Offers:

• Next-Gen AI: Using cutting edge techniques from academia and the commercial sector, AlphaK creates a unique network that combines static, contextual data with streaming spatiotemporal observations to derive knowledge and take action.
• Proven Accuracy: The AlphaK model has been tested with independently generated government evaluation data. The results were stellar—greater than 95% accuracy for real-time, streaming activity prediction with extremely low latency. The model also showed the ability to recommend productive actions when embedded in an RL framework.
• Transparent, Clean Visualizations: By leveraging internally-developed intuitive web-based visualization tool, called Agent Oracle, AlphaK provides situational awareness and clearly communicates. This tool allows playback of missions to while showing the key metrics needed to evaluate model performance and data integrity.
• Cloud-Optimized & Scalable: AlphaK includes a robust, modular architecture that is optimized for the cloud and efficiently leverages AWS autoscaling services to dynamically adjust resources based on mission demands. This optimized architecture helps customers meet timelines while minimizing cloud costs.

Automatically enhance EO and SAR satellite imagery and automatically extract actionable information.

MAGI provides a set of cloud-hosted AI applications that include:

• Data Labeler: Quickly and easily label targets of interest to greatly reduce the time to train and deploy AI models.
• Denoising Algorithm: Clean SAR imagery (removing speckle) when no clean ground truth is available to greatly increase interpretability and object detection algorithm performance.
• Spatial Resolution Enhancement Algorithm: Generate a finer resolution satellite image (8x upsampling) to improve aesthetics and enhance downstream AI performance.
• Object Detection Algorithms: Quickly produce bounding boxes to identify a diverse set of objects within remote sensing imagery.
• Instance Segmentation Algorithm: Identify unique objects and the pixels that comprise each instance to allow for automatic metadata generation, such as the tip-to-tip wingspan of an aircraft.
• Network Explainability: Help practitioners and users understand why the network made a certain prediction to build trust and increase the user adoption rate.

Automatically Analyzing Text, Categorizing Documents, and Managing Electronic Records to Meet NARA Directives.

Erudite is microservice-based solution that automatically analyzes documents and intelligently deciphers crucial metadata. Erudite determines which topic or category best aligns with a given document – in the machine learning field this is called document classification or categorization. These categories can be comprised from any records management framework, including the Bucket-Records Control Schedule (B-RCS) – the primary framework flowed down by the National Archives and Records Administration (NARA). Erudite also provides intuitive visualizations to allow records officers to understand what they manage, more easily create file plans, identify misplaced records, and support the incoming flood of new electronic records.

Erudite beat out three other vendors during a government bake off by demonstrating a high level of document categorization accuracy – achieving ~97% average accuracy on a benchmark dataset, while also being resilient to noise. It also demonstrated the ability to elastically scale in a cloud environment. Erudite was chosen as the most promising solution and our team was awarded a follow-on contract to deploy Erudite to the IC’s Commercial Cloud Services (C2S) – an example of Stratagem taking a program from concept to operations.

Erudite Offers:

• Traditional and Cutting-Edge: Erudite employs a unique combination of traditional and state-of-the-art technology. It combines a robust Latent Dirichlet Allocation (LDA) technique with cutting-edge models that top the General Language Understanding Evaluation (GLUE) benchmark index. The Erudite ensemble combines these complementary techniques to deliver greater accuracy, superior resiliency, and more secure operations compared to any single approach.
• Proven Accuracy: Erudite delivers 97% accuracy on the open source 20-newsgroup dataset. It also outperforms AWS Comprehend with the same data. This high level of accuracy led to the selection of Erudite for delivery into an operational environment.
• Does NOT Require Labelled Data: Our solution takes full advantage of any labelled data, but it is also able to support unlabeled data with unsupervised machine learning. Erudite generates “fingerprints” for documents and can analyze whether documents designated for the same B-RCS category are in the correct place.
• Records Management Dashboards: Erudite provides intuitive visualizations to help records managers organize and understand the proper disposition strategy for documents. With this web-based UI, it is easy to create or evaluate a file plan by highlighting the composition of records an officer needs to manage.