Multi-INT data fusion
Intelligence arrives as incompatible streams: imagery with geospatial metadata, signals detections with bearings and timestamps, radar tracks, and reports in free text. Fusion is the work of turning them into one model of the operating environment, with honest uncertainty attached to every element.
A fusion pipeline we build has five stages. Ingestion adapters normalise each source to a common schema: UTC time, WGS84 coordinates, error ellipses and source reliability. Association correlates observations to tracks using statistical gating on Mahalanobis distance, with global nearest neighbour for sparse scenes and joint probabilistic or multiple-hypothesis methods where targets are dense. State estimation runs Kalman and interacting multiple model filters for kinematics. Entity resolution links tracks, platforms, organisations and locations into a knowledge graph. Finally, an analyst workbench shows the result with full provenance: every fused fact points back to its source reports and keeps the source's reliability and credibility grading, in the Admiralty system style analysts already use.
The design rule is that fusion never hides disagreement. When two sources conflict, the picture shows both, weighted, rather than averaging them into a confident wrong answer.
ISR analytics: EO/IR, SAR and full-motion video
- EO/IR. Small-object detection at long range is dominated by a handful of pixels and atmospheric blur. We train on imagery at the operational ground sample distance, with thermal and visible bands modelled separately, and treat super-resolution outputs as cues rather than evidence.
- SAR. Synthetic aperture radar sees through cloud and darkness but brings speckle and layover. We use CFAR detectors for maritime targets, amplitude and coherent change detection between passes, and models trained on SAR directly rather than on optical imagery.
- Full-motion video. FMV delivered under STANAG 4609 carries KLV metadata (MISB ST 0601) with platform position, sensor pointing and frame corners. We georegister every detection per frame, so a track on screen is also a track on the map, and run multi-object tracking with stable IDs across occlusion.
Model outputs are cues for an analyst, not conclusions. Every alert carries a confidence score, the frame or chip that triggered it and the model version, and analyst confirmations flow back as training labels.
Detection and tracking at the edge
A full-motion video stream will not fit through a contested link. Running a quantised detector on a Jetson-class module on the platform, as described on our sovereign AI page, turns the stream into detections, image chips and track updates of a few kilobytes. The edge node keeps working when the link drops, buffers its tracks and forwards them in priority order when the link returns.
OSINT pipelines
Open-source intelligence pipelines collect from public news, social platforms, vessel AIS and aircraft ADS-B feeds, and commercial satellite imagery under a collection plan the customer approves. Multilingual NLP handles translation, named-entity recognition and event extraction. Near-duplicate detection collapses reposts into one item with its spread history, and content provenance signals such as C2PA content credentials are checked where present. Collection respects platform terms and privacy law, and the plan records what is collected and why.
C2 and the common operating picture
Command and control tools fail when they demand bandwidth the network does not have. Our common operating picture interfaces render MIL-STD-2525 or APP-6 symbology over offline map tiles, synchronise by delta rather than by full refresh, and exchange position and event data as Cursor on Target messages so they interoperate with TAK-based clients. Views are role-based: a user sees only the layers and sources their role is entitled to.
Communications in contested and denied environments
Tactical networks are intermittent by design and degraded by intent. We engineer for that:
- MANET. Mobile ad hoc networks form and heal without infrastructure, using proactive routing protocols such as OLSR. Throughput falls with each hop, so applications are built to send less, compress harder and prioritise.
- Software-defined radio. SDR platforms allow waveform agility, spectrum sensing to choose clear channels and rapid prototyping in GNU Radio before porting to fielded hardware.
- LoRa for low-rate telemetry. LoRa's chirp spread spectrum demodulates below the noise floor, and its short, infrequent bursts suit unattended sensors whose emissions should be brief. It is not a certified low-probability-of-intercept waveform, and we do not present it as one.
- SATCOM backhaul. LEO services give low latency and GEO services give fixed geometry. Both are exposed to jamming, so links are monitored and traffic is prioritised for degradation.
- Delay-tolerant networking. Store-and-forward with the Bundle Protocol (RFC 9171) moves data across links that are never simultaneously up end to end.
- EMCON-aware design. Devices support configurable emission control states: scheduled or commanded transmission windows, transmit power control, local buffering during silence and no unsolicited beacons.
Positioning, navigation and timing without GNSS
GNSS jamming and spoofing are routine in contested areas. We detect them by monitoring carrier-to-noise density, cross-checking constellations and testing receiver solutions for integrity, then fall back to other sources: tactical-grade inertial measurement, visual-inertial odometry, terrain-referenced navigation against elevation models, and timing holdover from disciplined oscillators. Our drones and robotics page covers the navigation estimators in detail.
Classified networks and cross-domain patterns
Systems at different classifications run on separate networks, and data moves between them only through accredited cross-domain solutions such as one-way data diodes and content-filtering guards, as set out in ASD's Fundamentals of Cross Domain Solutions. We design software to pass through them: schema-validated message formats with no executable content, protective markings carried with the data, offline installation bundles, no outbound calls and complete audit logs. Systems for government users are built against the Information Security Manual.
Technology on the Defence and Strategic Goods List is controlled under the Defence Trade Controls Act 2012. Amendments that commenced on 1 September 2024 created an AUKUS licence-free environment and new offences for supplying DSGL technology to non-exempt foreign persons, including inside Australia. We assess DSGL implications with you before technical data is shared. Security, clearance and accreditation requirements are agreed with each customer at the outset.
Questions
Who is this capability for?
Authorised defence, national security and government users, and the primes and integrators that supply them. Engagements are scoped to the customer's mission and authority.
Can the analytics run on a disconnected platform?
Yes. Detection, tracking and fusion components run on edge hardware without a network, buffer their outputs and synchronise when a link becomes available.
Which interoperability standards do you work with?
STANAG 4609 motion imagery with MISB KLV metadata, Cursor on Target for TAK interoperability, and MIL-STD-2525 or APP-6 symbology, plus customer-specific message formats.
Is LoRa suitable for military communications?
For low-rate unattended sensor telemetry it can be, because its bursts are short and infrequent. It is not a certified low-probability-of-intercept waveform, and voice or command traffic belongs on purpose-built tactical radios.