Production-ready edge AI platform that filters and prioritizes satellite imagery before downlink — saving 85%+ bandwidth and surfacing items of interest in real time. CNN + ReAct agentic reasoning + HMAC audit trail. Built for NVIDIA Jetson Orin, under 20 watts.
Or clone & run locally:
make install && make run
Real-time mission-control interface — pulls live Sentinel-1 SAR, Sentinel-2, and NOAA GOES imagery. Runs ship detection, fire monitoring, and agentic triage in your browser.
Above is a CSS-rendered preview of the live dashboard. The actual app pulls real satellite data and runs the same triage pipeline in your browser.
A complete onboard ISR stack — from sensor input to authenticated downlink decisions — with object detection, agentic reasoning, model versioning, and live satellite feeds built in.
MobileNetV3-Small at INT8 via TensorRT > ONNX > PyTorch with auto-fallback. Outputs cloud, anomaly, and value scores per tile.
Classical CFAR + connected-components vessel detector for Sentinel-1 SAR. Identifies "dark ships" — vessels broadcasting no AIS transponder.
Pure-Python think → act → observe loop activated on high-value tiles only. Under 2 W on Jetson. Optional Phi-3/Gemma SLM enhancement on AGX/Thor.
Every triage decision is logged with HMAC-SHA256 authentication and full provenance — tile hash, scene ID, scores, agent steps, timestamp.
Pulls real Sentinel-1 SAR, Sentinel-2 L2A, NOAA GOES-18, and NASA FIRMS imagery. No paid API keys required for the public sources.
SHA-256 checksums, training metadata, one-command rollback. Every deployed model is traceable to its training run and dataset.
Ground-station feedback loop: analyst CSVs or audit-log pseudo-labels → fine-tune → validate → auto-register if improved.
Live tegrastats integration on Jetson. Agent loop only runs when value score > 0.6 AND power budget allows. Hard ceiling at 20 W.
GitHub Actions: lint → test → security scan → SBOM → Docker + Trivy. Multi-arch Dockerfile, hardened with non-root user.
From raw sensor input to authenticated downlink decision in under 50 ms.
Tile-split, normalize
any spectral band layout
INT8 MobileNetV3
cloud · anomaly · value
YOLOv8 / SAR CFAR
items of interest
Agentic reasoning
only on high-value
KEEP or FILTER
with explanation
HMAC-SHA256
JSON Lines log
The platform is sensor-agnostic and mission-aware. The same engine runs wildfire monitoring, maritime ISR, disaster response, and pipeline inspection.
Real-time hotspot and smoke-plume prioritization. Cross-references with NASA FIRMS active-fire feed for ground truth.
Sentinel-1 SAR ship detection over contested waters. Flags vessels with no AIS broadcast — the actual intel signal.
Post-event damage assessment. Filters out unchanged tiles, surfaces collapsed structures and refugee movement.
Drone or satellite-based monitoring of pipelines, transmission lines, and offshore platforms for leaks and anomalies.
The API is intentionally minimal. Drop in any image array; get back a structured decision with full provenance.
The dashboard pulls real Sentinel-1 SAR over the Strait of Hormuz
and runs ship detection in your browser — clone & run in 60 seconds.
git clone && make install && make run
— full source under MIT license.