WISPAYR / AI ADVISORYStart →
AI/Casework/03 · iris-drone-detection
All casework
// CASE / 03 / security

Iris — drone detection at the edge

Ops teams running outdoor events & sensitive sites
securityeventsedgerf

RF + RTL-SDR + small ML on a Pi at the perimeter, alerting when a consumer drone shows up — no cloud round-trip.

// Cost
~£600 hardware per perimeter node. Build cost amortised across deployments.
// Duration
Iterating since 2025.
// 01 · The problem

Off-the-shelf drone detection is either £20k+ for a single sensor or relies on the manufacturer's own remote-ID broadcast (which jammers/grey-market drones don't carry).

// 02 · What we did

RTL-SDR sweeps known consumer-drone control bands, a small classifier identifies modulation patterns matching DJI/Autel/Skydio families. Pi + WireGuard back to the operations console gives sub-second alerting. Lives entirely on the customer's network.

// 03 · What the AI did

Spectral classification: is this RF signature consistent with a known drone control protocol?

// 04 · What humans did

Defined the alert criteria, integration with the venue's existing security workflow, and the call-the-police runbook.

// 05 · The outcome

Caught real drones at real events. False-positive rate manageable after one site-specific calibration day.

// 06 · What broke

Generic model trained on broad RF data was useless on the local noise floor. Per-site recalibration baseline is non-negotiable.

// 07 · What works

Hybrid sensing > any single modality. Now combining RF with audio and computer-vision confirmation for high-stakes deployments.

// 08 · Reusable lessons
  1. 01If your problem has expensive incumbents, the entry-price they leave on the table is your wedge.
  2. 02AI works best as the cheap broad sweep, with a human (or a second sensor) confirming.
  3. 03Edge deployment isn't just a privacy story — it's a latency and reliability story.