Layer 01Demonstrated
Passive optical sensing
USB camera input feeds a trained single-class drone detector running on NVIDIA Jetson Orin Nano edge hardware.
Platform // System architecture
From camera pixels to a coherent track.
Lamassu begins with passive optical observations at the edge. The platform direction is to correlate those observations across sensors, preserve their provenance, and produce useful track data for existing operator workflows.
Current focus // EO detection & tracking
System maturity // Operational prototype
The stack
Every layer has a job—and a status.
We distinguish what has been demonstrated from what is being developed. That boundary matters in a system intended for consequential environments.
Layer 02Demonstrated
Detection & visual tracking
The prototype classifies candidates, maintains visual identity over time, and can command a calibrated pan-and-tilt optical mount.
Layer 03In development
Observation-level fusion
A common observation model is intended to correlate optical, RF, radar, and acoustic inputs as those integrations mature.
Layer 04Roadmap
Air picture & C2 output
Structured track output for operator interfaces and command-and-control systems is planned, but is not currently available.
End-to-end direction
One architecture. Multiple useful signals.
The sensor does not need to know the whole air picture. It reports what it observed; the system decides how those observations relate.
Generate observations
Camera nodes create timestamped detections. Remote ID and other sensor adapters are planned as separate inputs.
Detect, filter, and track
Local inference reduces raw video movement, maintains visual tracks, and preserves sensor context.
Associate observations
Planned multi-camera triangulation and modality-agnostic fusion build confidence without hiding provenance.
Publish track data
The intended output is a coherent track that can flow into an operator view or standards-based C2 interface.
Current prototype
What works today.
Real-time camera inference
A trained YOLO detector runs against live camera input on Jetson edge hardware with an engineering HUD for timing and confidence.
DemonstratedVisual target persistence
ByteTrack-based tooling maintains target identity over video and exposes per-track motion information.
DemonstratedPhysical optical follow
Closed-loop proportional control drives a calibrated two-axis mount to keep the selected detection centered.
DemonstratedCross-node corroboration
A subscriber confirms alerts when multiple nodes observe activity within a time window, with deduplication and heartbeat monitoring.
DemonstratedFusion core
Confidence without losing the evidence.
The fusion core is in development. Its design preserves the sensors that contributed to a track, their health, and the confidence behind each update.
01Design principle
Observation first
Sensors report what they saw rather than independently declaring a complete track.
02Design principle
Provenance preserved
Every fused track is intended to retain which sensors contributed and how.
03Design principle
Health aware
Silent or degraded nodes should affect confidence instead of disappearing from operator context.
04Design principle
Pluggable sensors
Optical, RF, radar, and acoustic observations can enter through explicit adapters as integrations mature.
05Design principle
Replayable decisions
The same observation stream should produce the same track history for analysis and testing.
06Design principle
Open output
Track data is intended to support existing workflows rather than force a new closed ecosystem.
Evaluate the fit
Start with coverage, constraints, and the operator.
We will map your environment against the demonstrated prototype and make the maturity boundary explicit before proposing a next step.
Discuss your environment