They emit. We listen. That's all it takes.
A drone that cannot receive its pilot is not a drone — it is a crashing box of lithium. So every hostile quadcopter, every loitering munition, every FPV, every ISR platform in flight today emits a radio signal on a band we know. Nightjar is the silent, distributed sensor mesh that hears them, geolocates them to the metre, classifies them by model, and hands the track off to an effector — before the operator has finished his pre-flight checklist. Our sensors never transmit on the bands they watch. They cannot be found. They cannot be jammed. They cannot be spoofed by chaff, by noise, or by a well-written press release.
Small drones are not a solved problem. They are an opening act.
Every month of the current European security window reveals the same truth louder: plastic drones that cost less than a dinner are winning engagements against systems that cost more than a house. Four stress-lines are visible at once, and the old playbook does not cover any of them.
Radar does not see small drones. Not reliably. Not at useful ranges.
A 3 kg plastic quadcopter flying at 30 m above ground clutter has a radar cross-section comparable to a gull. Off-the-shelf surveillance radars designed for general aviation targets simply do not see it — or see it intermittently, at short range, after it has crossed the perimeter. Purpose-built counter-UAS radars exist and help, but they are expensive, limited in number, and themselves emitters. A loitering Shahed is a different class of problem; a 150 g FPV is a different problem again.
Jamming the whole spectrum is not a policy.
Wide-band RF jamming kills WiFi in the hospital, LTE in the passing ambulance, drone links belonging to the friendly reconnaissance platoon, GNSS for everyone in line-of-sight. It is legally fraught in peacetime, operationally dangerous in populated areas, and deeply counter-productive when you want to know which drone is there before you decide what to do about it. Blanket EW is a last resort dressed up as a first one.
An emitting sensor is a target.
A radar that transmits to find you will be found. Modern hostile drones — Lancet, Shahed derivatives, even commercial kit piloted by motivated operators — increasingly carry passive RF detectors of their own. An active counter-UAS radar installed at a base entrance is a priority target marked with a beacon. Europe is now deploying multi-million-euro radars that announce their own coordinates every second they are on.
Point solutions do not scale to contested skies.
One-off counter-UAS installations protect one airfield, one VIP, one power plant. Europe has thousands of airfields, borders, substations, refineries, depots, prisons, stadiums and ministries. The cost and footprint of single-site active systems make meaningful continental coverage financially and physically impossible. The only scalable primitive is a cheap, quiet, mesh-able passive sensor.
Silent. Precise. Distributed.
Three properties that in combination do not exist elsewhere. Silent because the sensor never emits on the bands it watches — zero counter-ELINT signature, nothing to bomb. Precise because hybrid TDOA + AOA geolocation drives position error below ten metres at useful ranges. Distributed because a Nightjar grid is a mesh of small cheap things — lose a node, the rest take up the slack; cover a continent one lamppost at a time.
Silent.
A Nightjar node is a receive-only system on the bands it monitors. No probing. No interrogation. No beacon. The RF front end is wired to antennas; the SDR samples IQ; the FPGA computes features; the mesh backhaul rides a separate physical layer (fiber, PoE, or non-overlapping radio). An attacker scanning the monitored bands finds nothing. Physically destroy one node and you learn the location of one node — nothing about the rest of the grid.
- Zero transmit duty on 400 MHz – 6 GHz monitored range
- Mesh backhaul on separate physical path (WiFi 6 / LTE / LoRa / fiber)
- Antenna array is indistinguishable on RF from a dumb receive dipole
- No GPS transmit, no LTE cellular modem required for operation
Precise.
One node hears a signal. Three nodes triangulate it. Four nodes locate it in three dimensions. Nightjar fuses Time Difference of Arrival across synchronised nodes with Angle of Arrival from phase-interferometric arrays, Doppler from range rate, RSSI from signal strength, and drone-model priors from the classifier — in an extended Kalman filter that outputs one honest track with one honest covariance. Meter-level accuracy is not a headline number, it is the boring steady state of a properly densified grid.
- TDOA timing resolution 10–50 ns → 3–15 m fix error
- AOA phase interferometry with 4-element array per node
- EKF fusion of TDOA + AOA + Doppler + RSSI + prior
- Honest covariance: the system tells you when it is unsure
Distributed.
Nightjar is designed mesh-first. Every node is equal. Every node talks to every other over Nexus Atlas multi-path (WiFi 6, LTE backhaul, LoRa fallback, fiber where available). There is no single fusion server you can knock out. There is no cloud the adversary can subpoena. A Nightjar cluster that has been shelled down to four surviving nodes is still a working counter-UAS system. A Nightjar grid that has lost its wide-area link is still a working air-gapped perimeter.
- No cloud. No mandatory central server. Fully offline-capable
- Per-cluster mesh fusion; wide-area aggregation is optional
- Graceful degradation — 3 nodes = 2D, 4+ = 3D, N-1 tolerance
- Nodes cost a fraction of an active radar. Continental coverage is arithmetic, not doctrine
Every radio a drone has to have — on bands it has to use.
A drone has to receive control and has to send video or telemetry to be useful. Both are radio. Both are on bands that are either legally mandated (Remote ID, ISM), physically convenient (5.8 GHz for low-latency video, 900 MHz for long-range control), or vendor-locked (DJI OcuSync 2.4/5.8). Nightjar covers all of them — the commercial hobbyist stack, the military FPV stack, the cellular-connected loitering stack, and the satellite-connected long-haul stack. One wideband receiver per node, tiled across the grid.
Band-by-band coverage — what is loud, what is quiet, what is rising.
FPV — analog & digital video.
5.8 GHz / 5.6 GHz / 2.4 GHzAnalog FPV broadcasts a continuous high-power video carrier — the loudest, easiest thing we see. Digital FPV (DJI O3/O4, HDZero, Walksnail Avatar) fingerprints by OFDM symbol rate, frame cadence and preamble. A strapped-on Shahed-style loitering munition and a racing quadcopter both leave a 5.8 GHz video breadcrumb the moment the goggles are on.
Control links — ELRS, Crossfire, FrSky.
868 / 915 MHz + 2.4 GHzThe modern open-source control stack has standardised on ExpressLRS at 868/915 MHz (long-range) and 2.4 GHz (low-latency). ELRS frequency-hops on a deterministic sequence and runs LoRa-class modulation; its preamble and hop pattern are a fingerprint. TBS Crossfire, FrSky R9, and Ghost carry their own signatures. Nightjar recognises all four families by modulation class alone — no decode required.
Commercial drones & Remote ID.
2.4 / 5.8 GHz + EU 2019/947DJI OcuSync / Lightbridge, Autel Skylink, Parrot Skycontroller, Skydio. Each vendor leaves OFDM burst patterns distinct enough to fingerprint model families. Remote ID broadcasts on 2.4 GHz WiFi/Bluetooth per EASA and FAA rules are also a passive signal — friendly drones with valid RID keys are filtered out cleanly; unregistered carriers are flagged immediately.
Cellular-tethered UAS.
sub-6 LTE / 5G NRA rising class of drones uses off-the-shelf 4G/5G modems to backhaul command and video over civilian networks — unbounded range, resilient against local EW. Nightjar detects their uplink PUSCH/PUCCH patterns and the drone's own modem hardware-ID signatures in RF. The carrier network need not cooperate for detection.
Satellite-connected drones.
Iridium / Starlink uplinkShahed-derivative long-haul loitering munitions and ISR platforms increasingly ride Iridium or Starlink uplinks. Their ground-to-sky transmissions leave a signature — Starlink user-terminal bursts at 14 GHz, Iridium L-band bursts at 1.6 GHz — that Nightjar's extended-band variants pick up cleanly from the ground.
Swarms & telemetry.
900 MHz / LoRa meshCoordinating swarms and loitering mesh architectures often use LoRa-class 900 MHz or proprietary FHSS mesh protocols. Telemetry back-haul on SiK radios at 433/915 MHz is the classic MAVLink signature. Both are narrowband, bursty, and deeply characteristic — easy to see, very hard to fake.
Protocol fingerprinting without decoding.
FHSS pattern.
Hop sequence, hop rate, dwell time. ELRS at 500 Hz hops differently from Crossfire at 150 Hz. No decode, just timing.
Preamble & sync-word.
Every radio puts known bits at the start of a frame to let the receiver lock. Those bits are a fingerprint. We see them in IQ without needing the key.
Symbol rate & bandwidth.
A 2.4 GHz DJI OcuSync burst is 20 MHz wide. A 2.4 GHz ELRS burst is 800 kHz wide. Occupied bandwidth alone separates half the zoo.
Duty cycle.
Analog FPV is a continuous 100 % carrier. DJI digital is 70–85 % depending on scene. ELRS is 0.5 % bursts. The ratio is a fingerprint.
Modulation class.
OFDM vs. FSK vs. LoRa chirp vs. analog FM. Distinguishable from cyclostationary features even at low SNR, even without a matched decoder.
Cyclostationary features.
Periodicities that survive spreading. For spread-spectrum and CDMA-like links, the cycle-domain reveals presence and type when the instantaneous spectrum looks like noise.
Phase-noise fingerprint.
Every consumer oscillator has manufacturing drift. Two DJI Mavic 3 units on the same firmware still transmit with different phase-noise tails. Hardware-level individuation, not just model.
Packet cadence.
Video streams pace frames. Telemetry paces heartbeats. ELRS paces channel updates. The time-between-bursts distribution is a per-protocol signature.
TDOA + AOA fused. A bearing, a hyperbola, a metre-scale fix.
Two complementary physical primitives. Time Difference of Arrival compares when the same RF burst hits two synchronised nodes — each pair defines a hyperbolic line of constant delay; three pairs intersect at the emitter. Angle of Arrival reads phase differences across a small antenna array on a single node to yield a direction. Neither alone is perfect. Fused in an extended Kalman filter, with Doppler from radial velocity and RSSI for sanity, they converge to sub-ten-metre accuracy in the steady state.
Three fixes, one truth.
TDOA · time difference of arrival.
HyperbolicThe same RF burst reaches Node A, B, C at times tA, tB, tC. The difference tA − tB is a constant along a hyperbola passing between A and B. Three hyperbolae from three node-pairs cross at a point — the emitter. Requires sub-microsecond time sync across all nodes (PTP over mesh, or GPSDO per node, or free-running disciplined clock as fallback).
- Timing resolution 10 – 50 ns drives position error 3 – 15 m
- 3 nodes → 2D fix · 4+ nodes → 3D fix · 5+ → overdetermined, robust
- Works on any burst that is heard at ≥ 3 nodes — narrowband or wideband
AOA · angle of arrival.
Phase interferometryFour antenna elements per node, each offset by a fraction of a wavelength. A wavefront arriving from a direction hits elements with known phase differences. Solve backwards (MUSIC, ESPRIT, or a simple phase-comparison correlator) to get bearing with sub-degree resolution. Two AOA bearings from two nodes cross at the emitter in 2D. Useful even when TDOA is impossible (single node on rooftop, for example).
- Array baseline λ / 2 per element · 0.5 – 3° bearing accuracy
- Independent of time sync — one node is sufficient
- Degrades with multipath and correlated arrivals; TDOA compensates
Doppler + RSSI · sanity.
Velocity & amplitudeMoving drones shift carrier frequency. Nightjar reads Doppler across multiple nodes to estimate radial velocity and, jointly with position rate, full 3D velocity. RSSI across nodes gives a coarse independent position prior — a passive cross-check on the TDOA/AOA math. The EKF will reject a "fix" that demands an impossible velocity or an RSSI geometry that cannot exist.
- Doppler resolution ≤ 1 Hz at 5.8 GHz → 0.05 m/s velocity
- RSSI triangulation as low-precision independent witness
- EKF covariance grows honestly when priors disagree; no false confidence
Node spacing drives range and precision. Pick your trade.
| Node spacing | Typical detection range | Fix accuracy | Altitude resolution | Deployment profile |
|---|---|---|---|---|
| 250 m | 3 – 6 km | 2 – 5 m | 4 – 10 m | Airport fence line · prison · VIP venue perimeter |
| 500 m | 5 – 10 km | 4 – 10 m | 8 – 20 m | Military base · refinery · high-value infrastructure |
| 1 km | 8 – 18 km | 8 – 20 m | 20 – 50 m | Forward operating base · large airfield · city centre |
| 2.5 km | 15 – 35 km | 20 – 60 m | 60 – 150 m | Border patrol line · regional early warning |
| 5 km | 25 – 50 km | 60 – 200 m | 150 – 500 m | Wide-area coverage · maritime flank · long frontier |
Design targets for a nominal 200 – 500 mW emitter at 868 MHz – 5.8 GHz, clear LOS from elevated nodes, στ ≈ 10 – 30 ns (GPSDO + 20 MHz cross-correlation).
Higher-Tx platforms (cellular-tethered UAS, SATCOM loitering munitions) extend detection range 2 – 5 ×. Accuracy degrades with GDOP when an emitter is far outside the cluster footprint; the EKF reports covariance honestly.
GPSDO per node.
Baseline: a GPS-disciplined oscillator locks each node's 10 MHz reference to GNSS. Timing accuracy < 30 ns 1σ in clear sky.
PTP over Nexus Atlas mesh.
When GNSS is denied or intermittent, nodes discipline each other via IEEE 1588 PTPv2 over the mesh. Master election is deterministic; holdover OCXO keeps sub-100 ns for tens of minutes.
Free-running fallback.
Total GNSS + mesh denial: nodes run on disciplined OCXO memory. Timing drifts predictably; the EKF tracks and reports degraded accuracy honestly, does not silently lie.
Not just "a drone." That drone.
Detection answers "is there a transmitter"? Classification answers "whose transmitter, how dangerous, how shall we respond"? Nightjar's classifier stack runs on each node's AI accelerator: per-band protocol class first, then model family, then specific model where the library has data, then friendly / hostile discrimination from signed Remote ID beacons where available. The classifier reports its confidence honestly. Unknown signatures are flagged for analyst review and become training data for the next site or the next rotation.
Commercial · hobbyist · military.
Class triageThe first useful cut. A DJI Mavic 3 looks nothing like a Shahed-136 ground-control uplink, which looks nothing like an ELRS-piloted FPV racer, which looks nothing like a Skydio X10. Class triage runs in < 50 ms from first burst and routes the track to the right operator surface: civil-compliance for Mavic, interception queue for FPV, immediate alert for Shahed.
Model fingerprinting.
Known libraryFor well-documented consumer platforms Nightjar resolves to specific model and firmware generation: DJI Mavic 3 Enterprise · OcuSync 3+, Autel EVO II · Skylink 2, Parrot Anafi USA, Skydio X10. The library is the product of a continual-learning pipeline, not a static list.
Adversary fingerprinting.
Shahed · Lancet · derivativeLoitering munitions and ISR platforms observed in Ukraine have begun to be characterised in the open-source spectrum-signatures community. Nightjar's adversary library grows with each encounter; we distribute updates out-of-band to sovereign customers. Shahed-family GCS uplinks, Lancet-family telemetry and FPV kamikaze kit are first-class entries.
Per-band classifier head.
One head per monitored band. Input: 50 – 500 ms of IQ slice. Output: softmax over protocol classes (OFDM-wide, OFDM-narrow, FHSS, LoRa, FSK, analog-FM, unknown). ~2 ms on-node.
Burst-pattern matcher.
Temporal CNN over preamble + first 8 bursts. Outputs model-family likelihoods. 5 – 10 ms. Anchors the expensive model lookup.
Library correlator.
Approximate nearest-neighbour match in a signed signature library. Unknown tail preserved for analyst review.
Remote-ID verifier.
Parses EU RID (EN 4709-002) and FAA RID beacons. Checks digital signature against trusted operator keys. Friendlies are filtered cleanly; unsigned or invalid beacons are flagged.
Continual learning — without phoning home.
On-site library growth.
Unknown emissions trigger a local archive entry: 500 ms of raw IQ, TDOA track context, classifier softmax with low confidence. An on-site analyst can label the archive at leisure; labelled examples re-train the on-node classifier head nightly with federated aggregation across the customer's own cluster. Nothing leaves the deployment unless the customer explicitly opts into a signed signature exchange with other Nexus Atlas sites — and even then the exchange is a signature fingerprint, not raw IQ.
Honest confidence, honest failure.
A classifier that always gives a best-guess answer in the face of noise is actively harmful — the operator starts to ignore it. Nightjar distinguishes confident classifications (> 85 %) from plausible (60 – 85 %) from ambiguous (< 60 %). Ambiguous tracks surface with a visual marker; the operator sees "unknown RF-emitting object" and can escalate rather than be told a wrong model family. Unknown tail is logged but never lied-over.
From drone power-on to operator decision in five seconds.
A drone must emit to fly. Nightjar hears the first packet, not the tenth minute. From the moment a hostile operator powers on a control link, the grid has the emitter localised, classified, tracked, pushed to the tactical map, and a response option staged — in the time it takes a person to read this sentence. Compare to a typical active counter-UAS radar: 15 – 60 s to classify a small UAS, if the cross-section is large enough to see at all.
Physical RF up, operator decision down. Six layers. One runtime.
Nightjar is a layered system with a single responsibility per layer and a replaceable implementation at every boundary. Apps depend on the fusion runtime, the fusion runtime depends on Nexus Atlas mesh, the mesh depends on the sensor abstraction, the sensor abstraction depends on the SDR / FPGA / GPSDO stack, and the physical RF layer is whatever antenna a mounting bracket can hold. Swap any layer for a sovereign equivalent without touching the layers above or below.
Replaceable at every boundary.
Every interface between layers is a versioned wire protocol. A sovereign customer can run the Nightjar firmware on their own SDR iron, a custom fusion runtime on their own server, a national classifier library on their own sovereign AI hardware, and we will honour it. We publish the protocol before we publish the product.
Where Nightjar ends, other systems begin.
Nightjar is a passive sensor — not an effector. Its job ends the moment a tracked track is handed, with confidence and covariance, to something that acts. We feed Compass natively, feed ATAK / SitaWare / ICC / AirC2BMC over standard tactical protocols via an integrator, and cue Phantom, Blackbird, or third-party kinetic and electronic-warfare effectors through the same pipeline. You are not locked in; you are made plug-in-able.
One box. Four antennas. Thirty watts. Everything else is software.
Nightjar One is the reference outdoor sensor node. A ruggedised IP67 enclosure, a four-element phase-coherent antenna array, a wideband SDR front-end, an FPGA for first-pass FFT and feature extraction, an ARM SoC with a 2 TOPS AI accelerator for per-band classification, a GPSDO for sub-microsecond time, and a Nexus Atlas mesh radio stack for talking to its neighbours. Pole-top, rooftop, vehicle-mounted, tripod, backpack — the same node in five form factors.
RF front end.
Layer 1Primary coverage 400 MHz – 6 GHz with a direct-conversion wideband receiver. Optional modular extension to 70 MHz – 18 GHz for HF/VHF and K-band. Four synchronous receive channels (one per array element). 61.44 Msps baseline; 200 Msps sprint. Sensitivity -95 dBm wideband, -120 dBm narrowband matched. 40 dB dynamic range, LNA + stepped-gain frontend.
- Primary 400 MHz – 6 GHz · extension 70 MHz – 18 GHz
- 4× synchronous IQ channels · 24 bit ADC equivalent
- Sensitivity -95 dBm WB · -120 dBm NB
- Bandpass filtering · LNA gain ladder · overload protection
Compute.
Edge fusionFPGA for FFT, burst trigger, decimation, feature vector extraction. ARM Cortex-A multicore SoC for drivers, classifier inference on a 2 TOPS AI accelerator (Hailo / Ambarella-class, vendor-agnostic), fusion math, mesh protocol, and local storage. On-board NVMe for 500 GB of IQ archive rolling buffer.
- FPGA · 524 288-point FFT, 120 MHz instantaneous BW
- ARM · 8 core · 2 TOPS AI accelerator
- RAM 8 GB · NVMe 500 GB
- Rust runtime · memory-safe · reproducible build
Time & reference.
Sub-μs syncGPS-disciplined oscillator as primary 10 MHz reference. OCXO holdover for GNSS-denied conditions — < 100 ns drift over tens of minutes. IEEE 1588 PTPv2 over mesh as secondary sync. 1 PPS distribution to FPGA for IQ sample timestamping. Every burst carries a timestamp good to < 30 ns in clear sky.
- GPSDO · 30 ns 1σ clear sky
- OCXO holdover · < 100 ns drift over 20 min
- PTPv2 over mesh · deterministic master election
- 1 PPS out · external-chassis sync support
Mesh & backhaul.
Nexus AtlasNexus Atlas multi-path stack: WiFi 6/6E for short-range high-throughput, sub-6 LTE / 5G modem for wide-area, LoRa for fallback low-rate mesh, fiber / PoE Ethernet where available. All four transports bonded through a single Noise IK + ML-KEM-768 tunnel. Aggregate throughput per node ≥ 100 Mbps across bonded links.
- WiFi 6E · 2.4 / 5 / 6 GHz · 800 Mbps peak
- LTE / 5G NR · 150 – 450 Mbps
- LoRa · 0.3 – 37 kbps fallback
- PoE++ Ethernet · fiber SFP · optional
Power.
Edge-powerPoE++ (802.3bt Type 4) from a single cable, or DC 12–48V for vehicle / tripod / backpack. Solar + lithium option for off-grid fence-line deployment — one 200 W panel and 1 kWh battery for 24/7 operation in sunny climates. ~30 W idle, ~60 W peak during heavy classifier inference. Designed for continuous duty cycle.
- Primary PoE++ · 90 W budget
- DC input 12 – 48 V
- Solar + battery option · 200 W panel + 1 kWh
- Draw · idle 30 W · peak 60 W
Environmental.
IP67 · MIL-810IP67 sealed aluminium enclosure. Operating range -30 °C to +60 °C. Designed to survive continuous outdoor exposure, Baltic winter, Mediterranean summer, desert dust, and salt-spray maritime deployments. MIL-STD-810 methods 501 / 502 / 507 / 509 targets; STANAG / formal qualification is an aspirational roadmap item, not a claim.
- IP67 sealed · aluminium + conductive gasket
- Operating -30 °C to +60 °C
- Humidity 5 – 95 % non-condensing
- Vibration / shock · design targets per MIL-810 guidance
Physical.
Mounting flexibility~4 kg with antenna array integrated. 310 × 240 × 90 mm main enclosure. Standard VESA-75 / pole-clamp / tripod / magnetic-mount adaptors. Four N-type ports exposed for external antennas where directional or elevated arrays are preferred. Tamper-evident case with detection into the fusion runtime.
- Mass ~4 kg · integrated array
- Dimensions 310 × 240 × 90 mm
- Mounts · VESA-75 · pole · tripod · magnetic · MOLLE-compat backpack
- Tamper sensor · reports to fusion runtime
Security.
Zero self-disclosureEvery node has a unique attestation key embedded at manufacture. Boot is measured; firmware is signed. All mesh comms are Noise IK + ML-KEM-768 hybrid encrypted. The node never transmits on monitored RF bands — the mesh runs on a physically separable transport, tuned to frequencies Nightjar is not watching. Seizing one node leaks one node's long-term keys; revocation propagates across the cluster in seconds.
- Measured boot · signed firmware
- Per-node attestation key · hardware root of trust
- Mesh transport on separable physical layer
- Compromise recovery · ratcheted per-cluster re-keying
Same silicon. Six field shapes. Ten minutes to stand up a grid.
The Nightjar node is the same box in every configuration. What changes is the mount, the power source, the antenna shape, and the concealment — not the firmware, not the mesh protocol, not the fusion logic. A customer that owns pole-top, vehicle, and backpack units can mix them freely; the cluster does not care where the antennas live.
Pole-top / rooftop fixed
The base case. Airport masts, rooftops, towers, substation lighting poles. PoE++ from a single Cat 6a cable up the pole; continuous duty; invisible unless you look for it. Typical 250 m – 1 km spacing.
Vehicle-mounted
On-the-move grid. Four nodes on four vehicles in convoy create a self-configuring cluster as they drive. Roof-rack array; vehicle 12 – 48 V power; automatic PTP sync over inter-vehicle WiFi 6E or LTE.
Tripod · rapid deploy
Two operators carry three nodes; three telescopic tripods up to 6 m; one click of inter-node mesh; one call to the fusion runtime. Full operational grid in under ten minutes from a vehicle tailgate.
Backpack kit
Each node fits a MOLLE-compatible pack with integrated battery (~8 h). Three packs carried by three soldiers is a 3-node grid. Deployment on foot across terrain inaccessible to vehicles.
UAV-borne / aerostat
Elevated sightline. A tethered drone or a small aerostat carries a node 50 – 300 m above ground. Horizon extends dramatically; a single elevated node reduces effective spacing required for the ground cluster.
Covert / buried
Low-profile variant without visible antenna array. Used for counter-surveillance missions where the grid's existence is itself sensitive. Drops into a utility pit, a vent, a fencepost. No RF signature to give it away.
What the adversary will try. What we already do about it. Where we are honest about limits.
Every counter-UAS marketing deck lies by omission. Nightjar will not. There is a real, narrow set of drones that Nightjar cannot see — and that gap is important to name, not hide. There is also a much larger set of evasive behaviours the adversary will try; we have answers for those that are rooted in physics, not marketing.
Frequency-hopping spread spectrum.
ELRS · Crossfire · DJIMost modern drones hop. ELRS cycles hundreds of channels per second across a band; DJI OcuSync hops OFDM subcarriers on a pattern; Crossfire runs a pseudo-random sequence. Nightjar does not chase the carrier — it tracks the modulation fingerprint across the hop set. The hop sequence itself is a fingerprint, and because the receiver covers the whole band at once, hopping is cosmetic to us. It confuses narrowband counter-UAS radios; it does not confuse Nightjar.
- Whole-band receive — 120 MHz instantaneous BW
- Cyclostationary feature detection — robust to hopping
- Hop pattern itself classified as signature feature
Bursty / low-duty-cycle links.
ELRS at 250 Hz · swarm meshVery short bursts with long silences are a classic stealth play. The longer the drone flies, the more bursts we hear — the first fix may be degraded; subsequent fixes sharpen rapidly. For a 250 Hz ELRS link, three bursts arrive within 12 ms; a TDOA solution is available before most humans have finished blinking. A single-burst-then-crash is beyond our honest window; we say so.
- First-burst fix possible above -110 dBm
- Three-burst confident fix within < 50 ms for 250 Hz links
- Degraded accuracy reported honestly, not hidden
Spread spectrum · CDMA-like.
low-SNR signaturesBelow the instantaneous noise floor by design. Presence is still detectable via cyclostationary feature analysis — the periodic pilots, preambles and frame boundaries are visible in the cycle-frequency domain even when the instantaneous spectrum looks like Gaussian noise. We detect presence and approximate location; we often do not decode payload. That is normally enough.
- Cycle-domain feature detection
- Presence / location without payload decode
- Works against well-designed LPI waveforms
Drones inside jammed / noisy spectrum.
local EWFriendly jamming raises the noise floor in the area being jammed. An FPV flying inside a jammed bubble is still emitting — the jammer is louder than the drone, but the drone's signature is not erased. Nightjar's per-band matched filters and high-dynamic-range front end still recover fingerprints; AOA degrades more than TDOA in heavy interference. We report this in the covariance.
- High-dynamic-range front-end · 80 dB IP3
- Matched filters survive in-band jamming within limits
- Honest covariance · track marked "degraded" when it is
Dense-carrier environments · urban.
WiFi / Bluetooth / LTE everywhereThe city is loud. 2.4 GHz is saturated with consumer WiFi and Bluetooth; 5.8 GHz is full of backhaul and outdoor APs; LTE is continuous. Naive RF detectors raise permanent alarms. Nightjar's classifier rejects known-friendly signatures aggressively (access-point MAC fingerprints, beaconing cadences, vendor OUI patterns in packet timing) and only surfaces novel RF consistent with drone-class emitter characteristics. False-positive rate on WiFi/Bluetooth is < 0.5 % per hour in our urban trials.
- Per-band noise-floor tracking and adaptation
- Known-carrier fingerprint library (WiFi, Bluetooth, LTE, FM, DVB-T)
- Surfaced alerts = novel, drone-consistent, classified
Dark drones — no uplink, inertial only.
the honest gapA drone that never receives and never transmits — pre-programmed waypoints, INS-only, no video downlink — cannot be heard. Nightjar does not see these. Examples: some loitering-munition terminal-phase configurations, certain cruise-missile-class platforms, fully autonomous inertial airframes. These platforms also cannot receive new tasking or abort commands, so their mission profiles are limited and their populations rare. The right answer is layered: pair Nightjar with Blackbird (EO / passive RF cue) and optical / acoustic sensors. We will not claim to solve this gap. We will tell you where the gap is and suggest what covers it.
- Honest constraint · pre-programmed inertial airframes are invisible to RF
- Mitigation · pair with optical / acoustic / radar for final layer
- Operational note · population of truly-dark drones is rare
A grid that sees. A map that knows. Effectors that act.
Nightjar is one sensor among siblings. Its value multiplies when it rides on Nexus Atlas transport, paints into Compass, cues Phantom, hands off to Blackbird, and shares an SDR substrate with Backtrack and Bastion. Same radios. Same mesh. Same keys. Same operator surface. A customer that deploys two of our products has a stronger position than a customer with both installed as point solutions — and that is by design.
The Rust daemon that bonds every available link (WiFi, LTE, LoRa, fiber) into a single Noise IK + ML-KEM-768 tunnel. Nightjar's mesh backhaul is Nexus Atlas, end-to-end. Time sync, telemetry, fused tracks — all flow through the same tunnel.
Nightjar tracks appear natively in Compass as first-class map entities — class, confidence, covariance, velocity, altitude, cue options. An operator on a Compass tablet sees the drone marker appear at T+1 s and can dispatch a response from the same surface.
When Nightjar classifies an inbound as ISR or GCS-teleoperated, Phantom is cued to the emitter's line of sight. The decoy swarm floods the drone's receiver with plausible false tracks — Nightjar provides the geometry Phantom needs to be convincing.
For high-value hostile drones — loitering munitions, ISR platforms — Blackbird is vectored to intercept from a Nightjar track. Nightjar's RF fix bootstraps Blackbird's own onboard passive RF and EO sensors. No active radar cue, no emissions to alert the target.
Shared SDR substrate. A Nightjar node in a border-site deployment can run Backtrack co-travel detection in time-slice — one hardware footprint, two missions. Useful for forward bases where payload is precious.
Nightjar's sub-6 LTE/5G coverage overlaps Bastion's IMSI-catcher / rogue-base-station detector. A deployed Nightjar cluster is also a Bastion cluster — detect drones, detect covert cellular surveillance vehicles, from the same hardware.
No self-disclosure. No mandatory cloud. No foreign kill switch.
Trust is not a slogan. It is a list of explicit, testable properties — each of which can be verified, broken, or honoured. Nightjar's trust architecture is five such properties, each pinned to a design decision, each measurable.
Zero self-disclosure on monitored bands.
Nightjar sensors do not transmit on 400 MHz – 6 GHz (or the extended 70 MHz – 18 GHz) range they monitor. The mesh backhaul runs on separable physical transports tuned to frequencies Nightjar does not watch. An adversary scanning the bands Nightjar is listening on finds nothing. Measurable: an in-band spectrum analyser placed next to a Nightjar node, scanning its monitored range, sees only ambient RF.
End-to-end encryption · hybrid post-quantum.
Every track, every alert, every telemetry packet leaves a node inside a Noise IK + ML-KEM-768 hybrid tunnel. Classical-and-post-quantum both must break to read it. Long-term node keys are pinned at manufacture and revocable per-cluster. Session keys ratchet. A seized node leaks only the session-window it held, not past or future traffic.
Sovereign jurisdiction.
Bulgarian HQ · EU data residency · source escrow available to sovereign customers. No US ITAR/EAR-controlled component in the critical path. No opaque phone-home update channel. Updates are offered; they are not pushed. A customer can run a Nightjar deployment forever on last-known-good firmware without asking anyone's permission.
Air-gap by default.
A Nightjar cluster is fully operational with no external network access. Cloud is a feature, not a requirement. Federated learning of new signatures, cross-customer library exchange, and remote operator support are all opt-in and signed. A customer that disables every outbound connection retains full detection, classification, and fusion capability.
GDPR-clean by construction.
Nightjar is a passive RF sensor. It does not retain non-drone-class signals. Friendly Remote ID data is processed under EASA / EU 2019/947 rules and anonymised once a friendly is resolved. WiFi / Bluetooth traffic heard incidentally as background is discarded at the feature-extractor stage — never stored. Site log retention is configurable per customer policy, default 30 days.
Export-control posture.
Nightjar is designed inside the EU and targets EU Common Military List Cat. ML11 and Cat. ML15 equivalents for the defence variant; dual-use Cat. 5A001 / 5D001 mapping for the civil-infrastructure variant. Formal classification is jurisdiction-specific and proceeds with the customer's national authority — this page does not claim a specific licence; it commits to transparent paperwork.
Design targets — real numbers, honest ranges.
Numbers on this page are design targets for the reference outdoor node. Concept stage; field numbers will land inside these envelopes when customer deployments begin. We do not claim specific formal certifications — STANAG, FIPS 140-3, EASA approvals are an aspirational roadmap, not a present-tense claim.
| Radio front end | |
|---|---|
| Primary frequency coverage | 400 MHz – 6 GHz · direct-conversion wideband receive |
| Extended coverage (optional) | 70 MHz – 18 GHz · modular HF / VHF / K-band front-end |
| Instantaneous bandwidth | 120 MHz baseline · 200 MHz sprint mode |
| Synchronous receive channels | 4 (one per antenna element) · phase-coherent |
| Sensitivity · wideband | -95 dBm at 20 MHz BW · noise figure < 3 dB |
| Sensitivity · narrowband matched | -120 dBm at 1 kHz BW · matched filter |
| Dynamic range · IP3 | 80 dB · stepped-gain LNA ladder |
| Sample rate | 61.44 Msps baseline · 200 Msps burst |
| Timing & synchronisation | |
| Time-sync method · primary | GPSDO · < 30 ns 1σ in clear sky |
| Time-sync method · fallback | IEEE 1588 PTPv2 over Nexus Atlas mesh |
| Holdover accuracy | < 100 ns over 20 min on OCXO · disciplined drift |
| TDOA timing resolution | 10 – 50 ns (drives 3 – 15 m position accuracy) |
| Detection & fusion | |
| Typical FPV detection range | 5 – 20 km at 1 km node spacing · 200 – 500 mW emitter |
| Typical commercial-UAS range | 8 – 30 km · OcuSync / Lightbridge / Skylink class |
| Typical cellular-UAS range | 20 – 40 km · sub-6 uplink bursts |
| Geolocation accuracy | 3 – 15 m at 1 km spacing · < 2 m at 250 m spacing |
| Altitude resolution | 5 – 40 m depending on node spacing & 3D geometry |
| First-track latency | 50 – 200 ms TDOA fix · 200 – 500 ms classified track |
| Classification confidence target | > 85 % confident on known library · ambiguous tail preserved honestly |
| False-positive rate | < 0.5 % per hour in urban spectrum (trial target) |
| Cluster & mesh | |
| Minimum cluster · 2D fix | 3 nodes |
| Minimum cluster · 3D fix | 4 nodes |
| Typical cluster size | 8 – 32 nodes · up to 64 per fusion runtime |
| Mesh throughput per node | 100 Mbps aggregate bonded · WiFi + LTE + LoRa |
| Mesh failover budget | < 800 ms path re-selection |
| Compute & storage | |
| FPGA · FFT / feature extraction | 524 288-point FFT · burst trigger · decimation |
| SoC · classifier & fusion | ARM 8-core · 2 TOPS AI accelerator (vendor-neutral) |
| RAM · NVMe | 8 GB RAM · 500 GB NVMe IQ archive |
| Power & environmental | |
| Power draw | 30 W idle · 60 W peak |
| Power input | PoE++ 802.3bt Type 4 · or DC 12 – 48 V · solar+battery optional |
| Weight | ~4 kg with integrated array |
| Dimensions | 310 × 240 × 90 mm |
| Enclosure | IP67 sealed aluminium · conductive gasket · tamper sensor |
| Operating temperature | -30 °C to +60 °C |
| MTBF target | > 50 000 h (design target · not yet field-validated) |
| Security | |
| Mesh encryption | Noise IK + ML-KEM-768 hybrid · ChaCha20-Poly1305 · X25519 |
| Boot posture | Measured boot · signed firmware · per-node attestation key |
| Compromise recovery | Cluster-wide re-keying within seconds of revocation |
| Log retention | Configurable · default 30 days on-node · GDPR-policy aware |
Two physics. Two postures. Two very different deployment stories.
Active radar is a mature, valuable, limited tool. Nightjar is a complementary primitive with a different shape — and, in most small-UAS scenarios the market actually cares about, a decisive advantage. This is an honest comparison on the dimensions that drive procurement.
| Dimension | Nightjar — passive RF grid | Active counter-UAS radar |
|---|---|---|
| Emissions posture | Zero transmit on monitored bands. Silent. | Primary emitter. Announces its own location continuously. |
| Small-UAS detection | Fires on the first control packet. Plastic airframe irrelevant — we read the radio. | Small RCS is physics. Purpose-built CUAS radars work; generic radars miss. |
| Classification | Fingerprint to protocol in < 500 ms. Model family where library has data. | Range, velocity, track — yes. Drone-vs-bird heuristics only; no native model ID. |
| First track latency | T+200 ms from drone power-on. | 15 – 60 s for small UAS · if ever detected at peak-RCS aspect. |
| Spectrum licence | None required. Receive-only. | Per-site licence · per-country · often slow. |
| Counter-attack exposure | No target for passive RF-seeking adversary drones. | Priority target · drones that home on emitting radars exist. |
| Jamming resistance | In-band jamming raises noise floor; classifier survives within limits. Track marked degraded. | Jammed radar is a blind radar. Full denial is a standard Russian playbook. |
| Scalability | Mesh-native. Continental coverage is arithmetic — node count × price. | Per-site deployment; high per-unit cost; finite coverage. |
| Dark-drone coverage | Honest gap. Non-emitting inertial drones are invisible to RF. Pair with optical / acoustic. | Sees any sufficiently-reflective airframe. Better here. |
| Friendly discrimination | Signed Remote ID parse · operator-key library · immediate friendly filter. | IFF integration varies; RF-based discrimination usually not native. |
| Urban deployment | Legal · quiet · fits a lamppost · does not jam civilian LTE. | Complex licensing · EMI-sensitive · spectrum-intense. |
| Per-node cost class | Low — fraction of a radar. Designed to be many. | High — designed to be few. |
Perimeters, borders, and events — wherever drones are a threat and spectrum is crowded.
Eight deployment patterns drawn from current European and allied procurement conversations. Each drives different trade-offs in node density, form factor, and fusion policy — all run the same firmware.
Airport perimeter
A ring of rooftop / pole-top nodes around the fence line. Passive, license-free, compatible with busy civilian spectrum. Remote ID verifier cleanly passes commercial delivery drones; unregistered incursions trigger a runway-closure alert to ATC in under one second.
Military base
Zero-emissions posture is the entire point. An active radar at the gate is a target marker. A Nightjar ring is invisible, mesh-surviving, and runs indefinitely on PoE++. Hand-off to organic air defence and deployed EW kit over standard tactical protocols.
Land border · drone incursion
Lines of masts along the eastern flank detect loitering ISR drones crossing from Kaliningrad, Belarus, Crimea, Black Sea. 2.5 – 5 km spacing gives wide-area early warning; cue to organic fighter / SAM batteries over Link 16 gateway (integrator-supplied).
Maritime flank
Vessel-mounted nodes network over WiFi 6E + SATCOM to shore clusters. Detect Shahed-class inbound, FPV kamikaze from small boats, ISR over international water. The maritime deployment pairs with onboard EW and kinetic effectors.
Critical infrastructure
Contested perimeter protection. Low-profile nodes on lighting masts and facility rooftops. Integrates into the site SCADA / physical-security console via native or CoT. Ammunition depots and chemical-plants are a priority deployment pattern.
VIP events & political summits
Rapid-deploy backpack kit. Two operators stand up a 5-node tripod cluster in a plaza in ten minutes. Compass tablet in the protection detail's hands. Unregistered drones overhead are pinned and dispatched before the principal arrives.
Prison perimeter
Contraband delivery by small drones is a daily operational problem across European corrections. Nightjar ring detects the operator powering the link outside the wall; site security intercepts before the payload drop. FPV and commercial platforms both covered.
Forward operating base
Deployed expeditionary kit. MANET / SATCOM mesh, solar-battery autonomy, backpack-deployable tripod nodes around the FOB perimeter. Integrates into coalition C2 over CoT or NFFI. Runs for weeks on last-known-good firmware without phoning home.
The honest questions. The honest answers.
Questions we hear in every serious customer conversation. If something important is missing here, write to office@NexusAtlas.io.
What about dark drones — the pre-programmed inertial-only kind that never emit?
What happens when GPS is denied — how do you synchronise?
Sybil attacks — can an adversary flood the classifier with fake emissions?
What is the false-positive rate on WiFi and Bluetooth in urban spectrum?
< 0.5 % per hour per node. Nightjar maintains a per-band noise-floor model and a known-carrier fingerprint library — WiFi access-point beacon cadence, Bluetooth Low Energy advertising patterns, LTE downlink reference signals, FM broadcast, DVB-T — all learnt and suppressed. Surfaced alerts are novel RF events consistent with drone-class emitter characteristics (modulation, duty cycle, bandwidth, hop pattern). A single WiFi hotspot appearing on a new channel does not trigger an alert; a 500 mW 5.8 GHz analog carrier most certainly does.
Is it legal to deploy Nightjar in urban areas? What regulatory burden?
What export-control class applies?
What is the minimum viable deployment? How small can a cluster be?
Does weather matter? Rain, fog, snow?
0.1 – 0.3 dB / km attenuation — detectable in the link budget, not a showstopper. Fog is essentially transparent at our monitored bands. Snow accumulation on antennas is a mechanical concern (heated radomes in polar-climate variants); the receiver itself is unaffected. Extreme solar / ionospheric events can perturb HF-band detections on the extended-coverage variant, but not the primary 400 MHz – 6 GHz window. Nightjar runs through Baltic winter and Mediterranean summer without behavioural change. Weather is much less of a problem for us than for EO or IR sensors — this is a systemic Nightjar advantage in bad-weather theatres.
Can Nightjar detect drones at very low altitude — NOE / tree-top?
What about drone swarms — 50+ simultaneous emitters in a small area?
> 64 concurrent tracks per fusion runtime with stable identity. Larger swarm configurations cue denser clusters or a second fusion runtime.
How does Nightjar differ from the counter-UAS RF systems already on the market?
Nightjar is a concept. Your problem is not.
Nightjar is early stage — TRL 1–2. There is no hardware to ship and no demo to book. What there is: a design, a portfolio it plugs into, and a team that wants to talk to the people whose fence line, airfield, refinery, or forward base is currently uncovered by anything that works against a small drone. If that is you — or if you fund the people it is — start a conversation. We will tell you honestly what the product is today, what it will be in twelve months, and whether your timeline is realistic.
Two ways in.
Operator briefing — for end-users, integrators, and procurement teams evaluating passive RF detection as part of a layered counter-UAS posture. We walk the architecture, the honest limits, the deployment math, and the integration path.
Technical briefing — for engineers, research groups, and prime contractors. We go deeper: classifier stack, fusion runtime, mesh transport, clock architecture, SDR front-end choices, and where we are looking for partners.