2026-04-02
Low-Power Inference for Embedded ISR: A Field Report
Eighteen months of deploying inference engines on constrained platforms has taught us where the power budget actually goes — and it's not where most neural architecture papers assume.
TECHNICAL INSIGHTS
Engineering notes from the field.
Technical writing from the Kestrelsense engineering team. Topics cover latency-constrained neural inference, sensor fusion architecture, hardware SWaP tradeoffs, mmWave radar in contested environments, and human-on-the-loop design. Written for systems engineers and program technical leads — assumes working knowledge of embedded compute and unmanned platforms.
2026-02-26
Human-on-the-Loop: Design Patterns for Supervised Autonomy
The phrase 'human in the loop' is overloaded and underspecified. We distinguish four distinct supervisory patterns and discuss when each is appropriate for autonomous ISR platforms.
2026-01-22
A Practical Architecture for Multi-Modal Sensor Fusion on Autonomous Platforms
Fusing EO/IR, radar, and acoustic sensor streams on a resource-constrained platform requires careful timestamp alignment, cross-modal uncertainty weighting, and graceful degradation.
2025-12-18
SWaP Tradeoffs in Edge AI Hardware Selection
Jetson Orin, Hailo-8, Myriad X, custom FPGA — every edge compute platform involves a different point on the latency / power / cost curve. We document the tradeoff space for mission-critical ISR payloads.
2025-11-12
mmWave Radar in Contested RF Environments: Design Lessons
mmWave radar offers sub-centimeter resolution in low-visibility conditions, but contested electromagnetic environments introduce interference patterns that most commercial radar stacks don't anticipate.
2025-10-15
Latency-Constrained Neural Inference for UAS Payloads
When an unmanned aircraft has 40 milliseconds to classify a target and can't call home, every compute cycle matters. We walk through our approach to sub-50ms inference under real SWAP constraints.