DroneWatch AI: Building a Multi-Sensor Drone Detection Demonstrator
Earlier this year I entered DroneWatch AI in the EUDIS Defence Hackathon in Bucharest. The idea stayed with me, and I have now built it into a working demonstrator. This is what it is, how it works, and why a software layer, not more hardware, is the interesting part of drone detection.
First, honestly: this is a concept demonstrator, not an operational system. The drone-threat signals are simulated. The air-traffic registry is a live, real ADS-B feed. It does not detect real drones. I built it to prove the approach and to start conversations with partners.
The problem: cheap drones, expensive detection
Commercial drones now threaten airports, critical infrastructure, public events and borders. The usual answer is specialised counter-drone hardware, which is expensive, so coverage stays sparse and hard to scale. Meanwhile a single sensor is noisy: radio-frequency alone, acoustics alone or radar alone each throw false alarms, and operators get flooded with unfused, unexplained detections they cannot act on.
The idea: fuse weak signals, do not buy more hardware
Plenty of cheap, distributed sensors already exist. The value is not another sensor. It is the software intelligence layer that makes the ones you have smarter, correlates them, and hands a clear, explainable track to whoever is on watch.
DroneWatch fuses five sources into one picture per zone:
| Source | What it contributes |
|---|---|
| Radio-frequency | The drone's control and video link |
| Acoustic | The rotor signature |
| Telemetry | Position and Remote-ID style data |
| Registry / ADS-B | Which flights are authorised (this feed is live and real) |
| Human reports | What a person on the ground sees |
Built to not cry wolf
The most important design choice is false-alarm suppression. A single sensor blip never alerts. A registered, authorised flight, even a loud one, stays quiet. A HIGH assessment requires at least two independent sensors to agree and a contact with no matching authorised flight. An alert that is wrong most of the time gets ignored, and in this domain that is the failure that matters.
Live data, a tactical map, and an AI that explains itself
The demonstrator plots live, real air traffic from a public ADS-B feed on a tactical map, around a chosen protected site, next to the simulated drone track. A two-tier engine assesses each situation: deterministic rules first, then an AI layer that explains, in plain operator language, why a given threat level is justified by the signals present. It runs on infrastructure I own and monitor, the same way I build and run every other custom AI system for founders.
Built to integrate
Being usable means speaking the interfaces operators already run. The design ingests real sensors and pushes a scored track into an existing command system through recognised standards:
| Domain | Interfaces |
|---|---|
| Identification | Remote ID (ASTM F3411), ADS-B |
| Air and drone traffic | ASTERIX, U-space / UTM |
| Military command and control | STANAG 4586 / 4609, Link 16 |
It stays strictly defensive: detection and early-warning only, with any response left to a separate, authorised, human-in-the-loop system.
Where it goes next, and who I am looking for
The honest next step is a pilot at one real site with one or two real sensor types plus the live feeds, to measure the real-world false-alarm rate, the number that actually matters. From there it grows toward machine-learning detection and a distributed grid.
I am looking for serious partners to build this into a fielded system: counter-UAS and sensor companies, systems integrators, and people working on European defence innovation. If you bring the sensors and the accreditation path, I bring the AI fusion layer. My background is as an Army Engineer (Military Technical Academy, Bucharest), so I understand both the operational world and the software. If that is a fit, I would value a conversation. You can also read what a paid AI pilot looks like or hire a fractional AI engineer.
Frequently asked questions
Is DroneWatch AI an operational counter-drone system?
No. It is a concept demonstrator. The drone-threat signals are simulated and it does not detect real drones. The air-traffic registry it cross-references is a live, real ADS-B feed. It exists to prove the fusion approach and to start partner conversations.
How does it reduce false alarms?
It requires agreement. A single sensor blip never raises an alert, and an authorised, registered flight never does. A HIGH assessment needs at least two independent sensors to agree and a contact with no matching authorised flight.
Does it replace radar or RF sensors?
No. It is the software fusion and AI layer above whatever sensors you already have. It makes distributed, mixed sensors smarter and hands a scored, explainable track to your command system.
What would it take to field it?
Real sensor integration at one site, a pilot to measure the real false-alarm rate, and a partner who brings sensors and an accreditation path. EUDIS and European Defence Fund calls fund exactly this kind of dual-use software layer.
Who built it?
Marius Andronie, an Army Engineer who now builds and operates production AI systems. DroneWatch AI began as an entry to the EUDIS Defence Hackathon in Bucharest.
This project was first entered in the EUDIS Defence Hackathon Spring 2026, Romania, part of the EU Defence Innovation Scheme.
📚 Related Resources
Get the 45-Point Acquisition Diligence Checklist
The complete pre-close checklist search funds, independent sponsors, and micro-PE buyers use to verify a business before they sign, free, and yours in one click.
Get the free checklist →