Case Study | Building Intelligent Cold-Chain Monitoring for Reliable Real-Time Alerts

Building Intelligent Cold-Chain Monitoring for Reliable Real-Time Alerts

Data Engineering Machine Learning IoT Integration

FiftyFive developed a cold-chain monitoring platform for the client, combining data engineering, IoT integrations, BigQuery-based data modelling, machine learning, backend APIs, and real-time alerting. We also built trailer-to-sensor mapping and an operations dashboard to improve data accuracy, alert reliability, and traceability.

Build Your Platform With Us

The Client

Client overview

The client operates refrigerated (reefer) trailers that move temperature-sensitive freight. Its trailers carry Samsara sensors that report temperature every minute, while trips and shipments are managed in the Ditat transport management system (TMS). The client needed to know, for every load, whether the product stayed within its required temperature range, and to be alerted quickly when it did not. The existing alerts fired so often on harmless events that operators had stopped trusting them.

The Need

Challenges

Reliable cold-chain monitoring meant joining two systems that were never designed to talk to each other, at minute-level volume, while making alerts trustworthy enough for operators to act on. Every reading also had to stand up to regulatory audit.

Sensor telemetry and trip data lived in separate systems. Samsara reported readings per sensor, while Ditat recorded trips, legs and shipments. Without a reliable join, a temperature reading could not be tied to the load it was protecting.
Sensors report at irregular intervals, and trips have hard start and end boundaries. Readings had to be aligned to trip time windows and gaps filled consistently, across multi-leg trips, mid-route trailer swaps and door-opening cycles.
Sensors were moved between trailers during hardware swaps and re-installations. Readings were silently attributed to the wrong trailer, creating data-integrity errors that no alert could catch.
Reefer units warm up during normal defrost cycles, cooling cycles and loading. Simple threshold rules treated all of these as breaches, so most alerts were false positives and genuine excursions were lost in the noise.
Operations teams needed alerts in the tools they already used, with priority levels and without repeated notifications for the same event.
Refrigerated freight is regulated, so every reading had to be traceable to its trip, shipment and trailer for audit.

Our Approach

Solution

FiftyFive built the platform on Google BigQuery as a layered data model that turns raw sensor readings into trip-aware, shipment-attributed records. A dedicated trailer-to-sensor mapping service keeps attribution correct through hardware changes. BigQuery ML classifiers then decide whether each temperature excursion is a genuine breach or a normal operating cycle. FastAPI services on Cloud Run and Cloud Functions serve the results to a React control centre and Slack.

Telemetry ingestion and medallion data model

FiftyFive built the ingestion and modelling layer that makes every temperature reading trip-aware and auditable.

Ingestion of more than 1.4 million Samsara sensor readings a day at one-minute granularity.
Integration with the Ditat TMS API for trip, leg and shipment metadata.
A bronze, silver and gold (medallion) model in BigQuery, separating raw, cleaned and business-ready data.
Time-window alignment and interval gap-filling to reconcile irregular sensor reporting with trip event boundaries.
Business rules for multi-leg trips, mid-route trailer swaps and door-opening cycles.

“Phone Book” trailer-to-sensor resolution

FiftyFive engineered a mapping service that keeps every reading attached to the right trailer.

A Cloud SQL (MySQL) registry of which sensor is installed on which trailer, and when.
Correct attribution through hardware swaps and re-installations.
Removal of a class of silent errors that had been assigning readings to the wrong trailer.
An API for viewing and maintaining trailer mappings.

ML root-cause classification

FiftyFive trained machine learning models to tell real breaches apart from normal reefer behaviour.

BigQuery ML classification models that attribute the root cause of each temperature excursion.
Separation of genuine cold-chain breaches from defrost, cooling and loading cycles.
Continuous retraining as new trip and sensor data arrives.
Training and scoring inside BigQuery, without moving data to a separate ML platform.

Alerting, APIs and operations control centre

FiftyFive put the platform into production with alerting tuned for the operations team.

Priority thresholds so the most serious excursions surface first.
Suppression and cooldown rules that stop repeated alerts for the same event.
FastAPI services on Cloud Run and Cloud Functions for trip data, alert history and trailer mappings.
A React control centre for operators to review trips, alerts and mappings.
Slack notifications that send alerts to the channels the team already uses.

Technology

Tech Stack

Data Platform

Google BigQueryMedallion Data Model

Machine Learning

BigQuery MLRoot-Cause Classification

Cloud & Backend

Google CloudFastAPICloud RunCloud Functions

Mapping Registry

Cloud SQLMySQL

Telemetry & Integrations

Samsara SensorsDitat TMS APISlack

Frontend

React

Experts

Team

AI & Data
Engineers

The Impact

Results

Project DurationJuly 2025 – Sept 2025

FiftyFive delivered a cold-chain platform that ties every reefer temperature reading to its trip, shipment and trailer, and alerts operators only when a breach is real. Operators regained trust in alerts, and compliance teams have a complete, auditable temperature record for every load.

70% fewer false-positive alerts

ML root-cause classification cut false-positive temperature alerts by 70%, so operators act on alerts again instead of ignoring them.

Trip-aware data at scale

More than 1.4 million readings a day are aligned to trips, legs and shipments at one-minute granularity.

Correct trailer attribution

The Phone Book service keeps readings on the right trailer through sensor swaps and re-installations.

Audit-ready cold chain

Every reading is traceable to its trip, shipment and trailer, improving regulatory compliance for refrigerated freight.

Support

FAQs

How does the cold-chain monitoring platform work?+

The platform joins Samsara temperature readings with trip, leg and shipment data from Ditat TMS in BigQuery. A trailer-to-sensor mapping service keeps attribution correct, while BigQuery ML classifies excursions and FastAPI services deliver results to a React control centre and Slack.

How did machine learning reduce false-positive alerts?+

BigQuery ML classifiers distinguish genuine cold-chain breaches from normal defrost, cooling and loading cycles. This root-cause classification reduced false-positive temperature alerts by 70% in the client’s platform.

How much sensor data does the platform process?+

The platform ingests more than 1.4 million Samsara sensor readings a day at one-minute granularity and aligns them to trips, legs and shipments.

How are Samsara readings linked to Ditat trips?+

The platform integrates Samsara telemetry with the Ditat TMS API for trip, leg and shipment metadata. Time-window alignment and interval gap-filling reconcile irregular sensor reporting with trip start and end boundaries.

What is the Phone Book trailer-to-sensor service?+

The Phone Book is a Cloud SQL (MySQL) registry that records which sensor is installed on which trailer, and when. It keeps readings attributed to the correct trailer through hardware swaps and re-installations, with an API for viewing and maintaining mappings.

Why use a medallion data model in BigQuery?+

The bronze, silver and gold layers separate raw, cleaned and business-ready data. This layered model turns raw sensor readings into trip-aware, shipment-attributed records that are auditable.

How does the platform handle trailer swaps and irregular readings?+

Time-window alignment and interval gap-filling reconcile irregular readings with trip boundaries. Business rules account for multi-leg trips, mid-route trailer swaps and door-opening cycles, while the mapping registry tracks sensor installations.

How are repeated temperature alerts prevented?+

Priority thresholds surface the most serious excursions first. Suppression and cooldown rules stop repeated notifications for the same event, and Slack sends alerts to the channels the operations team already uses.

Where are machine learning models trained and scored?+

BigQuery ML classifiers are trained and scored inside BigQuery, without moving data to a separate ML platform. The models are continuously retrained as new trip and sensor data arrives.

What can operators review in the control centre?+

The React control centre lets operators review trips, alerts and trailer mappings. FastAPI services on Cloud Run and Cloud Functions provide trip data, alert history and mapping information.

How does the platform support cold-chain audits?+

Every temperature reading is traceable to its trip, shipment and trailer. The platform provides a complete, auditable temperature record for every load to support regulatory compliance for refrigerated freight.

What team and timeline delivered this platform?+

Three AI & Data Engineers delivered the platform from July 2025 to September 2025 for a logistics and transportation client in the USA.