/case-studies/moonsyst

agritech · iot + ml 2023

Moonsyst.

Smart rumen bolus for dairy and beef farmers.

An IoT sensor that lives in a cow's rumen for years, plus an ML platform that turns biosignal noise into hours-accurate predictions of oestrus and calving.

python scikit-learn pytorch react-native fastapi aws iot

At a glance

// the client

A livestock-tech company building sensor hardware and software for dairy and beef farmers across Europe and the Americas.

// the challenge

Turning raw rumen biosignal into decision-grade predictions accurate enough to be the system of record for breeding and birthing.

// the outcome

Hours-accurate predictions of oestrus and calving, deployed across 1,200+ farms. Now the operational standard for the company.

The problem

What wasn't working.

[ 01 / 06 ]

Existing livestock monitoring depended on visual observation, breeding records, and external sensors that lost signal in barn environments. Farmers were missing oestrus windows and being surprised by calving — both expensive failures. Moonsyst had built the sensor; we were asked to build the ML platform that made the sensor decision-grade. Success looked like predictions farmers could plan a Tuesday around — accurate to within hours, calibrated, surfaced before the event.

Approach

What we did.

[ 02 / 06 ]

edge-light, cloud-heavy: The bolus does minimum work; the platform does the rest. Real-time on-device inference is cost-prohibitive for a sensor that lives in a rumen for years.
weekly retraining: Models retrain on aggregated farm data with per-farm signature isolation. No cow gets confused with another cow.
time-aware models: Handle sensor drift, calibration windows, and the long sparse signal patterns characteristic of bovine biology.
human-in-the-loop: The platform suggests; farmers decide. We didn't automate decisions a vet should make.
// where AI helped: Scaffolding, integration code, synthetic test data generation, model evaluation harness. Feature selection and sensor noise handling stayed engineer-led.

What we built

Four components.

[ 03 / 06 ]

01 · ml-pipeline

ML pipeline

Temperature, pH, motion, and rumination feature extraction. Oestrus and calving classifiers with calibrated confidence.

python · scikit-learn · pytorch

02 · backend

Backend

Sensor ingestion, per-animal state, alerts pipeline, integrations with third-party farm-management systems.

fastapi · postgresql · clickhouse

03 · mobile

Mobile

Farmer-facing alerts, herd overview, calving timeline. Designed for one-handed use in barn conditions.

react-native · ios · android

04 · cloud

Cloud & ops

Batch retraining, real-time inference, observability across 1,200+ deployed sensors.

aws · docker · kubernetes

Outcomes

What it shipped as.

[ 04 / 06 ]

97%

Oestrus detection within a 4-hour window

92%

Calving prediction within a 6-hour window

1,200+

Farms deployed across 14 countries

18mo

In production with <0.5% false alert rate

"Moonsyst is now the system our breeding manager checks before he opens his eyes in the morning."
— farm operator · ireland

The team

Who shipped it.

Four senior engineers: ML lead, backend lead, mobile lead, DevOps. ML and mobile leads from Lumio; backend and DevOps hybrid with the client team.

engagement: dedicated team · 11 months to production · ongoing retainer

What we'd do differently

Honest reflections.

Stronger investment in sensor-side observability earlier. Some early debugging time was spent in the cloud when the root cause was a sensor calibration drift we should have surfaced from day one.

Related work