Creating Mission Focused AI Solutions
Customer Success Stories
Practical AI for Multi-Physics Sensor Platforms
The unseen problem with multi-physics and multi-sensor AI inference is the excessive cost of producing labeled data. We encountered this problem with a commercial sensor platform doing remote condition monitoring for industrial and residential applications. To overcome this, we minimized our dependence on supervised machine learning and data labeling by using self-supervised and semi-supervised learning models. We adapted our culture, process, and tools to accommodate auto regressive techniques rather than relying on data labeling. This significantly reduced our staffing from 100s of people to under a dozen and minimized cost, timeline, and risk. We achieved super-human level of performance in ambient scene recognition and remote condition monitoring. Using our SpeedShift™ approach we fielded the system in 6 months.
26% Data Quality Improvement
Information from different sources found in most data warehouses contain conflicts that are hard to find and resolve. Differentiating between similar people, places, and organizations is essential to accurate and actionable intelligence. Common approaches for deconflicting data use static rules for choosing an authoritative representation. Our novel use of knowledge graph relationships allows us to develop ML models that identify and correct erroneous, conflicting, and deficient information. In a recent project data quality increased by 26% based on federal standards.