Defense
Defense technology
Local-first AI infrastructure engineered toward NIST 800-53 and IL5 control objectives (no formal assessment completed). A standalone pure-Go safety engine for any LLM. A distributed AI mesh for contested environments. Counter-UAS interceptors with autonomous engagement guidance. Sub-2M-parameter detection across six domains. Open-source Rust autopilot. CBRN sensor fusion. Arabic NLP. 27 USPTO provisional patent applications across 10 patent families — all filed March 2026, sole inventor, sole assignee, no government funding.
Compliance
Engineered toward NIST 800-53 / IL5 controls
Every control below is implemented in source — but no formal security assessment, third-party audit, or accreditation has been completed. These are engineering design targets, not certified compliance claims.
HMAC-chained tamper-evident JSONL audit log; server refuses to start without the audit key (AU-9)
Trusted CA chain, CRL checking, cert pinning, EDIPI extraction — implemented in source, not yet validated for deployment (IA-2(12), SC-17)
FIPS 140-3 build mode with runtime self-test via CMVP-validated module (product-level validation not yet initiated) (SC-13)
Five levels (UNCLASSIFIED → TS) plus CUI designations (NOFORN, IMCON, PROPIN, ORCON); constructor refuses by design to disable (AC-4)
Unicode normalization, lookalike-character detection, entropy-based secret detection (>4.5 bits), fuzzy redaction (IR-9)
DoD 5220.22-M three-pass overwrite (0x00 / 0xFF / random) on sensitive files (MP-6)
Default-on; bcrypt cost 12; 12-character minimum password; rate limiter 5 attempts / 5 min (IA-2(1))
Code-interpreter sandbox network policy by classification: UNCLASSIFIED → bridge, CUI → none, SECRET / TS → denied; fail-closed on unknown (SC-7, AC-4)
3,363 test functions, 18 fuzz tests in the security package alone, classification-routing fuzz harness gated against bypass
Featured program
DroneBane — three tiers, one kill chain
A counter-UAS product family covering the entire engagement workflow: a passive Remote-ID app on the operator's phone (Watch), a commercial sensor-fusion + MAVLink C2 platform for force protection (Pro), and a ATAK plugin designed for defense use with 16-point safety interlocks and on-device DRL guidance (Mobile). 8 USPTO provisionals across the family. Anduril Lattice integration prepped.
DroneBane Watch
DetectPassive Remote-ID, multi-sensor fusion, FAA reporting
DroneBane Pro
Engage (Roadmap)MAVLink C2, fleet management, autonomous patrol
DroneBane Mobile
AuthorizeATAK plugin, 16-point interlocks, F2T2EA, DRL guidance
Programs
All 13 active programs
Each card lists the program's USPTO provisional docket numbers (THRN-001 through THRN-027), TRL, and live demo link where available. Patent counts are non-overlapping except where explicitly noted.
DroneBane
Three-Tier Counter-UAS Product Family DetailsA three-tier counter-UAS family covering the full kill chain. DroneBane Watch is a passive Remote-ID + sensor-fusion app for force protection and base security. DroneBane Pro adds MAVLink C2, fleet management, and autonomous patrol (roadmap) for force protection and FOB defense. DroneBane Mobile is the ATAK plugin designed for defense use: 16-point safety interlocks, F2T2EA kill chain, on-device DRL guidance, FIPS 140-2 cryptography, Anduril Lattice integration prepped. See the full deep-dive on the dedicated DroneBane page.
RigRun
Local-First AI Inference Stack — Engineered Toward NIST 800-53Local-first AI infrastructure engineered toward NIST 800-53 and IL5 control objectives. Classification-gated routing across UNCLASSIFIED → TS with anti-bypass spillage detection, FIPS 140-3 build mode with runtime self-test, HMAC-chained tamper-evident audit logging, DoD 5220.22-M three-pass secure delete, and a purpose-built Docker sandbox with classification-aware network policy. CAC/PIV client authentication is implemented in source but not yet validated for deployment. No formal security assessment, audit, or accreditation has been completed. 51,000 lines of security code, 3,363 test functions, 18 fuzz tests.
HawkStack
Sub-2M-Parameter Detection Across 6 DomainsTopology-driven model architecture that predicts the performance ceiling of a dataset before training, then builds models under 2M parameters that beat 30–60M-parameter SOTA competitors. Validated across thermal drone detection (ThermalHawk, 82.95% mAP, 1.77MB INT8), IRST / missile warning (SentryHawk, 92.29% IoU), sonar target detection (DepthHawk, 88.0% mAP), ECG arrhythmia classification (CardioHawk), PCB defect inspection (ForgeHawk, 97.6% mAP), and wildlife camera trap ID (WildHawk). One architecture, six domains, edge-deployable on microcontrollers and SBCs (Jetson, Kria, COTS-ruggedized).
Pyros
Pure-Go LLM Safety EngineStandalone 17,000-line Go safety engine that wraps any LLM in a 7-pillar pipeline. Selective-buffer streaming safety proxy intercepts only tool-use blocks for evaluation — zero added latency on text streaming. PID homeostasis controller degrades gracefully under resource contention instead of failing. Drops in front of RigRun, OpenAI, Anthropic, vLLM, llama.cpp, or Ollama. No Python, no CGO, ships as one binary.
Mycelium
Distributed AI MeshDistributed AI mesh that turns commodity devices into a self-healing inference cluster over consumer WiFi. Three-layer proof-of-inference protocol — behavioral fingerprinting, Merkle execution traces, asymmetric reputation — is designed to verify untrusted nodes without trusted hardware. Classification-aware routing keeps UNCLASSIFIED traffic anywhere, CUI on the mesh, SECRET on the originating device. Built for DoD contested comms, disaster response, and GPS-denied environments. 40-claim provisional.
Meridian
Open-Source Rust AutopilotFlight controller autopilot written from scratch in Rust. 47 crates covering attitude control, navigation, sensor fusion, and mission planning. Open source. Distilled from ArduPilot with functional parity targeted through systematic functional testing.
CAIRN
CBRN Sensor Fusion DemoReal-time contamination zone estimation from heterogeneous CBRN sensor networks. Spatial grid-accelerated fusion with per-sensor anomaly detection. ISA 2.0 canonical data model with embedded IDLH tables and CAIRN-extended Cursor-on-Target XML namespace.
Agent Factory
Autonomous Meta-Agent Manufacturing14-step pipeline that designs, adversarially hardens, and deploys domain-expert AI agents from plain-English specifications. Requirements decomposition, multi-pass prompt synthesis with independent critique, automatic tool selection from a typed registry, knowledge base assembly, adversarial stress testing across 5 probe types, and a human approval quality gate. Six agents currently in use including the SBIR Proposal Writer used for federal grant work.
Aurora
Acoustic Swarm Communications DemoCovert data transmission via engineered rotor acoustics. GPS-denied cooperative positioning through propeller-modulated acoustic ranging. Sub-decimeter accuracy, zero RF emissions. Shares blade-geometry acoustic patent (014) with the DroneBane platform.
Harakat
Arabic DiacritizationPost-processing system for Arabic text diacritization. Contrastive dual-head neural corrector with adaptive multi-zone confidence thresholding. Hierarchical triple-key error-report disambiguation. 0.68% diacritization error rate, 6.7MB model, 62× smaller than SOTA at comparable accuracy.
Learned DCLAHE
Neural Image EnhancementDifferentiable Contrast Limited Adaptive Histogram Equalization with backpropagation-optimized parameters. End-to-end trainable image enhancement module that plugs into any detection pipeline (including HawkStack) for degraded visual environments.
Canopy
Multi-Modal C-UAS Situational AwarenessMulti-modal counter-UAS situational awareness platform combining EO/IR/RF/acoustic sensor inputs through Dempster-Shafer evidence fusion. 8-layer detection-classification-tracking-prediction-engagement stack designed to integrate DroneBane, ThermalHawk, and Mycelium into a unified C-UAS picture.
TRAVERSE
S&T Decision Support DemoTechnology reconnaissance and venture ecosystem research platform. Investment tracking, technology radar, MCDA scoring, SBIR pipeline analysis. 21 NIST controls implemented.