Hardened edge inference engineered for unmanned platforms
Sub-15ms classification. Under 8W. 95 grams. Designed for MIL-SPEC deployment from the ground up.
Classification latency is a mission-critical constraint, not a performance preference
Defense systems integrators, autonomous-vehicle OEMs, and program offices procuring ISR payload and edge-AI compute for unmanned platforms face a capability gap that existing hardware cannot close. Autonomous systems deployed in contested environments require sub-100ms sensor-fusion and threat classification at the edge — but the available perception stacks force a choice that should not exist.
Capable hardware — GPU clusters and full-rack inference servers — delivers the accuracy but cannot fit in a Group-2 UAV payload bay. It consumes 25–80W, requires cloud offload for classification, and introduces more than 200ms of round-trip latency. During that 200ms, the platform has moved. The target has moved. The engagement window has closed.
Compact solutions that fit the form factor consume so little compute that they cannot run multi-modal fusion. Single-sensor EO classification misses targets behind camouflage and thermal cover. The result is a platform that can carry a sensor but cannot act on what the sensor sees — not because the physics are impossible, but because the hardware has never been built for this specific constraint set.
Latency exceeds engagement windows
State-of-practice edge inference solutions require more than 200ms round-trip to cloud classification. Most ISR engagement windows are 80–150ms. The math does not work with cloud dependency.
Power budgets exclude capable hardware
Group-2 platforms allocate 5–15W total payload power. Existing capable inference boards consume 25–80W. The only available options are either too power-hungry to carry or too limited to classify.
RF-denied CONOPS eliminate cloud fallback
Contested environments involve deliberate RF jamming and GPS denial. Any perception stack that relies on cloud connectivity fails the moment it enters the operational environment it was purchased to address.
Commercial boards lack MIL-SPEC hardening
Adapted commercial AI boards expose attack surfaces through unsigned firmware, non-validated boot chains, and thermal design limits that fail below -20°C. Defense payloads require cryptographic root of trust, not retrofit security patches.
From raw sensor stream to cueing vector in under 15ms
Multi-modal sensor ingestion
The KS-100 module attaches to EO/IR gimbals and LIDAR payloads via standard MIL-STD-1553 or ARINC-818 interfaces. Raw sensor streams — simultaneous EO imagery, LWIR frames, and LIDAR point clouds — feed the on-module neural inference engine in real time. No preprocessing delay, no buffering to a host computer first. The module begins inference from the first available frame.
On-module neural inference
A custom NPU running quantized transformer models performs multi-modal sensor fusion — correlating EO, IR, and LIDAR returns to classify objects, estimate kinematics, and generate cueing vectors at 60fps with no cloud dependency. The inference pipeline is deterministic: maximum latency is bounded at under 15ms regardless of scene complexity. The NPU draws 6–8W peak, within the thermal budget of a Group-2 payload bay operating at -40°C to +85°C.
Timestamped target tracks and cueing data
Timestamped target tracks with classification confidence, bounding volumes, and cueing coordinates are streamed to the host vehicle's mission computer over encrypted serial or Ethernet. The module also outputs compressed metadata logs in STANAG 4609 format for post-mission ISR analysis. DDS and ROS 2 middleware output is available for integration with modern autonomous vehicle stacks including PX4 autopilot and L3Harris Falcon Edge.
Six capabilities engineered as a complete system
Each KS-100 design decision — power, thermal, security, form factor — is optimized for MIL-SPEC deployment, not retrofitted to it.
Sub-15ms Inference Engine
The KS-100's custom NPU executes 6-class object detection and kinematic estimation at 60fps with end-to-end latency under 15 milliseconds — measured from raw pixel read-out to cueing vector output. The inference pipeline is deterministic: maximum latency is bounded regardless of scene complexity, a property required by hard-real-time mission computers. All processing is on-module; the system remains fully operational during RF-denied or GPS-denied flight.
Multi-Modal Sensor Fusion
Operating from a single EO camera alone leaves classification vulnerable to camouflage and thermal crossover. The KS-100 ingests simultaneous EO, LWIR, and LIDAR point-cloud streams and runs a cross-modal attention model that resolves ambiguities a single-sensor approach cannot. The fused output is a labeled 3D bounding volume with classification confidence and velocity estimate — reducing false-positive alert rates versus single-modal baselines.
8W Power Envelope
Group-2 unmanned platforms allocate 5–15W total payload power; legacy edge-AI boards consume most of that budget before the sensor even powers on. The KS-100 draws 6–8W peak under full inference load, leaving power headroom for comms and actuators. The module operates across -40°C to +85°C without active cooling — designed for CONOPS where the payload cannot carry a fan.
Hardened Secure Boot Chain
Defense payloads face adversarial firmware replacement as a mission-degradation attack vector. Every Kestrelsense module ships with a hardware root of trust that validates each stage of the boot chain before execution. Inference model updates are signed with program-specific ECDSA keys; unsigned or modified models are rejected before any sensor data is processed. Hardware attestation is available for integration with program-level supply-chain verification frameworks.
SWAP-C Optimized Form Factor
Kestrelsense KS-100 measures 68mm x 40mm x 12mm and weighs 95 grams including connectors. A hardened conformal-coat variant is available for open-air pod integration. The mechanical interface follows standard MIL-DTL-38999 connector assignments, matching the footprint of common EO payload control boards so vehicle integrators can substitute the KS-100 without redesigning the bay or rerouting harnesses.
Field-Updateable Model Store
Operational requirements change faster than hardware cycles. The KS-100 maintains a model partition separate from the OS partition; authorized mission planners can push new inference graphs to the module over the avionics bus or ground-link within minutes. Model swap completes without a reboot — enabling mission-profile adaptation between sorties. The model store holds up to four independent inference graphs, allowing rapid switching between detection, tracking, and change-detection profiles at task-order speed.
Built for defense integrators evaluating edge-AI payloads — not for every platform
Primary Segment
Defense prime integrators and Tier-2 suppliers building ISR payloads for Group-2 and Group-3 unmanned platforms. DoD program offices evaluating edge-AI for contested-environment ISR under PEO Aviation or PEO IEW&S. SBIR/STTR Phase II programs with payload budgets in the $5M–$200M range seeking pre-production qualification partners.
Program Context
Programs that require hardened avionics bus integration (MIL-STD-1553, ARINC-818), ITAR-controlled firmware delivery, and formal SBIR subcontract or engineering service agreement structures. Integrators who need interface documentation and qualification data before initiating a procurement conversation, not after.
Not a Fit For
Commercial UAS logistics and delivery platforms, consumer drones, cloud-connected smart-city cameras, agricultural imagery platforms, or any system that tolerates classification latency above 100ms. If your CONOPS assumes a persistent RF link and no adversarial electronic environment, KS-100's specific design constraints are not the right trade-off for your program.
Engagement Process
Engineering sample requests begin with a technical inquiry via the contact form. We provide interface documentation, power and thermal characterization data, and a preliminary integration assessment before any contractual commitment. Evaluation samples are available under standard SBIR subcontract or engineering services agreement structures.
Standard avionics bus and middleware compatibility
Request engineering documentation
Interface specifications, power characterization data, and integration guides are available for defense integrators under technical evaluation.
Contact Engineering Team