$42M in Cargo Saved. 250K Shipments Monitored. 15-Minute Intervention Window.

The architecture starts at the sensor. Every refrigeration unit, shipping container, and transport vehicle was fitted with IoT temperature sensors transmitting via MQTT every 15 seconds: temperature, humidity, GPS coordinates, battery level, and sensor health. MQTT handles the connectivity challenges gracefully. A sensor on a refrigerated truck crossing rural Montana does not have reliable LTE. MQTT queues messages when connectivity drops and delivers them when it returns.

The data flows into two parallel paths. Path one: real-time alerting. Sidekiq Enterprise workers evaluate readings against compound threshold rules. Not just "is this above 8 degrees" but "has this sensor reported three consecutive readings above 7.5 degrees, with a rising trend, in a unit containing biologics with a 15-minute TOT window?" The rules engine is configurable by the client's quality team without engineering involvement.

Path two: TimescaleDB for time-series storage. Every reading, regardless of alert status, is stored for compliance records and analytics. TimescaleDB's automatic partitioning (hypertables) and built-in compression reduced query times to under 500ms and cut storage costs by 90% compared to vanilla PostgreSQL (which hit a performance wall at month three, taking 30+ seconds for historical queries).

The alert cascade: second 0, sensor reports breach. Second 15, second reading confirms trend (filters sensor noise). Second 30, Sidekiq triggers the alert. Second 45, ActionCable pushes WebSocket notification to dashboard. Simultaneously, SMS to facility manager and automated Twilio voice call to emergency response number. Second 60, the dashboard shows: affected unit location (PostGIS renders the facility map with the unit highlighted), current temperature, rate of change (degrees per minute), unit contents (product name, lot numbers, quantity, dollar value), and estimated time until product compromise.

The facility manager hears: "Unit 47 is at 9.2 degrees and rising at 0.3 degrees per minute. It contains $4.2 million in adalimumab. You have approximately 12 minutes. Nearest backup unit is Unit 23 in Building C." That specificity is the difference between a $4.2 million loss and a $0 loss.

FDA 21 CFR Part 11 compliance shaped every architectural decision. Append-only PostgreSQL tables with database-level triggers preventing UPDATE and DELETE on compliance records. SHA-256 hash chains linking records (if anyone modifies a historical record, the chain breaks and the system flags it). Two-factor electronic signatures for quality sign-offs. Over 400 RSpec tests serving double duty as both the regression suite and the basis for IQ/OQ/PQ validation documentation.

The predictive layer was unplanned. Six months in, analysis of billions of stored readings revealed that catastrophic compressor failures were preceded by subtle telemetry patterns: wider temperature oscillation cycles, longer compressor run times, gradual 0.1 to 0.2 degree drift over days. A gradient-boosted tree model trained on 18 months of data now flags units likely to fail within 72 hours. Facility teams receive weekly reports. Even with an 18% false positive rate, the early warnings reduced unplanned failures by 34%.

Heartbeat monitoring was a critical reliability feature. If a sensor misses two consecutive transmissions (30 seconds of silence), the system generates a connectivity alert. Operations teams learned that connectivity alerts at 2 AM often preceded temperature alerts at 2:15 AM because the same power failure that knocked out the sensor also knocked out the refrigeration unit.