N:SYM builds NeuroSymbolic AI systems that combine neural learning with symbolic reasoning to deliver explainable, deterministic decision-making for defense and security operations. Its BeliefNet core powers models like Edge and Observe, which perform real-time pattern recognition and object identification with plain-language justifications for every prediction. The platform runs on low-power portable devices, with inference speeds of approximately 1 millisecond per instance.
Funding
Funding not disclosed
Founders
Product
Problem
Traditional statistical AI models excel at pattern recognition but lack reasoning, explainability, and determinism, making them unreliable for high-stakes operational environments. Symbolic AI alone cannot adapt to new data, leaving a gap for systems that must both learn from experience and justify their decisions with traceable logic.
Solution
N:SYM provides a NeuroSymbolic AI platform that integrates neural learning with symbolic reasoning, enabling systems to know what is true, learn what is likely, and reason about what follows. Its core algorithm, BeliefNet, encodes expert human logic directly into neural networks, producing bounded reasoning engines that deliver deterministic, explainable predictions. The platform includes models such as Edge for through-life identity tracking, Observe for component-level object recognition, and Act for goal-driven autonomous planning, all running on low-power portable devices. Every inference is accompanied by a plain-language explanation traced back to the specific evidence that drove the outcome, ensuring operators can audit and trust the system's decisions.
Target Audience
Primary customers are defense and security organizations, including military operators, intelligence agencies, and surveillance teams that require reliable, explainable AI for mission-critical decision-making and autonomous systems.
Features
- BeliefNet core architecture split into a safety-critical inference engine and a flexible experimentation layer, ensuring consistent behavior in deployment while allowing rapid model development
- Explainability engine that answers analyst-style questions such as "what inputs mattered" and "what would change the outcome" in plain language with evidence traceability
- Edge model performs multi-modal fingerprinting using facial, shoe, apparel, and gait recognition, re-analyzing subjects every 5 frames to maintain identity across appearance changes
- Observe model learns specific objects from as few as 48 data samples and detects them even when they differ from training data, with component-level reasoning
- Act model plans over long time horizons while accounting for partially observable states and intrinsic/extrinsic goals, avoiding rigid rule-based failure modes
- Automatic model cards and trust scoring for every inference, plus novelty detection to flag inputs outside the model's experience
- Typical saved model size of 40–800KB and per-instance inference time of approximately 1ms, enabling deployment on low-power edge devices