Whisker Wardn

Non-lethal, welfare-first rodent deterrence — driven by computer vision

It spots rodents on camera and gently moves them on: no traps, no poison, no killing. A vision model watches the feed, and a signal that never repeats itself keeps rats from settling in.

Built during the Futurekind Spring Fellowship 2026

The problem, and a kinder answer

The usual response to rodents is lethal: snap traps, glue boards, anticoagulant poisons — cruel, damaging to the wider food chain, and ineffective because they never address why rodents keep arriving.

Whisker Wardn takes the opposite stance, built on a simple order of priorities:

01

Seal first

Physical exclusion is the only proven, permanent defence. The app says so plainly.

02

Watch

A vision model monitors any camera you already have — a fixed IP cam or an old phone — and recognises a rat instantly.

03

Deter, don't harm

On a sighting, a varied ultrasonic signal makes the space unwelcoming. Never lethal, never a tone the animal can learn to ignore.

04

Prove it

Every appearance and response is measured, so you can actually tell whether it's working.

The computer-vision model

The detector is the technical heart of the project — an object-detection model that runs entirely in your browser, so camera frames never leave the device.

ArchitectureUltralytics YOLO26 (nano) object detector
TaskSingle-class detection — Rat
ExportONNX, end2end=True — box selection runs inside the graph, no post-processing needed
Input[1, 3, 640, 640] float32 RGB, letterboxed, normalised 0–1
RuntimeONNX Runtime Web (WASM) — client-side inference, frames stay on the device

Running the model in the browser (rather than the cloud) means the whole detect→deter loop works on ordinary hardware — and no video ever leaves the premises unless you choose to send it to help improve the model. That privacy matters when the “camera” is a phone pointed at someone’s home or storeroom.

Why the signal never repeats

Cheap ultrasonic repellers fail because animals habituate — they learn a constant tone is harmless. Whisker Wardn is designed specifically to defeat that.

Every burst is different. Within the configured range (default 20–50 kHz, up to a 90 dB ceiling, ~5 s bursts), each pulse sweeps a randomised sub-band, in a randomised direction, with jittered length and cadence.

It learns what works — with transparent statistics, not a black box. A per-zone multi-armed bandit (softmax with a fixed exploration floor) picks each episode's frequency sub-band, loudness, sweep direction and emitter — biased toward whatever made rodents flee fastest, while guaranteed exploration keeps the signal unpredictable.

Welfare is enforced in code, not left to config:

45s
max sustained emission per episode
20s
cooldown before an emitter refires
12/hr
trigger ceiling per zone

Does it actually work?

A deterrent you can't measure is just a hopeful guess. Efficacy is a first-class feature.

Visits & episodes

Every appearance logged with confidence & duration; every response logged with its full signal profile.

Flight latency

How quickly the animal left after the signal started.

Return detection

Watches for up to two minutes after a burst — a quick flee that immediately returns isn't counted as success.

A/B baseline

The deterrent auto-cycles on and off in timed blocks, comparing activity against a true no-deterrence baseline at the same site.

Grounded in evidence

Every claim is tied to real, hand-verified literature — no invented citations.

Technical overview

YOLO26 · ONNX ONNX Runtime Web React + Vite Leaflet / OpenStreetMap Firebase Web Audio Netlify

A single-page web app: cameras (real IP cameras or a repurposed phone) feed the in-browser detector, detections drive the deterrence engine, and everything — device map, detection log, analytics, A/B efficacy — lives in one console. A pluggable hardware bridge drives a physical ultrasonic emitter when one is attached.

Credits & contribution

Whisker Wardn was created during the Futurekind Spring Fellowship 2026 — with thanks for the support, mentorship, and space to build a humane answer to a problem almost always solved with cruelty.

I am very grateful for the supervision and guidance of Vatsal Mehra — Cohort Facilitator.

The idea & humane vision

A welfare-first, non-lethal, exclusion-first approach to rodents.

Product flow & experience

How detection, deterrence and measurement fit into one usable console.

The research

Sourcing and verifying the scientific evidence the whole approach rests on.

The vision model

Designing, training and deploying the in-browser YOLO26 rat detector.

This app is primarily centered on computer-vision detection that I am consistently improving upon. The user interface and backend framework, however, were built rapidly with AI-assisted tooling.