BACKWATER

AI-Powered Insurgent Network Identification from Communications Intelligence

98% BALANCED ACCURACY | 0.00% CIVILIAN FALSE POSITIVE RATE | 4.09M RECORDS PROCESSED

Background:

During Actions on the Objective for a direct-action mission, individuals found on target will get questioned to determine their identity and or roll within a network. The team supporting this task will typically be briefed ahead of time with a series of questions to ask. After the information is gathered, any electronic device is typically collected and brought back for exploitation. What if that could be done in minutes on the objective?

Our Solution: Quantum-Inspired Advanced Neural Network

BACKWATER's three-stage architecture fuses advanced message passing, quantum-inspired superposition layers, and quantum attention across 9 feature groups — geospatial, app usage, communication metadata, network topology, temporal behavior, linguistic indicators, financial/OpSec, quantum embeddings, and keyword context. When testing synthetically generated data, Agentic Agents learn that context determines meaning, classifying each device as commander, operative, facilitator, asset, or civilian with explainable confidence scores suitable for direct F3EAD integration. These roles were created for synthetic data only and can be modified. Intelligence Analysts can train Agentic Agents to carry out specific and transparent tasks supporting human-machine teaming.

Why This Matters

  • Our testing received a 98% balanced accuracy with zero civilian false positives across 10,000 synthetic devices

  • Identifies 287 insurgent devices with no flagged SMS content — classified purely from behavioral patterns

  • Under 2 MB model footprint, full pipeline in under 7 minutes

Built for the Edge: Runs on the Overmountain Scoured Sentry (OSS)

Completely offline, solar-powered, under 23 pounds. Purpose-built for Actions on the Objective. Operators feed seized devices directly into the system at the point of collection and receive information about a device within a network. No reachback. No DOMEX queue. The F3EAD cycle on site.

Synthetic Data Only. All 4,093,399 records are synthetic, precisely mirroring Cellebrite UFED extraction formats from open-source documentation — no reformatting required. Operational validation with real-world data is required before any deployment consideration. Use our contact form to get in touch with us for additional information.