Supply Chain & Cold Chain

The Food Loss Problem

The US loses an estimated 30-40% of its food supply across the chain (USDA ERS), with fresh fruits and vegetables among the highest-loss categories. The economic cost runs to roughly $160 billion annually at retail and consumer level. The loss is not uniformly distributed: it spikes at three nodes where temperature, time, and handling quality break down simultaneously — at the farm gate (lack of precooling), at distribution-center cross-docking (mechanical damage + delays), and at the retail last mile (inadequate display refrigeration).

AI supply chain models attack all three nodes, but the highest ROI is at the distribution center and retail, where data infrastructure already exists (EDI/ASN transaction records, USDA AMS market reporting, 3PL GPS tracking) and where small interventions produce large loss reductions.

Open data/post-harvest-loss-model.csv — it contains loss-rate observations across 18 commodities (tomato, onion, banana, mango, potato, leafy greens, grapes, pomegranate) for 200+ supply chain legs, linked to temperature excursion frequency, transit time, handling score (assessed by field agents), and shipping origin. The target: predict loss rate (% weight loss + quality downgrade rate) for a given shipment configuration before dispatch.

Cold Chain IoT: Sensor Architecture and Anomaly Detection

A modern US cold chain instrumentation stack for a refrigerated truck (e.g., a Sysco or US Foods distribution reefer):

Open data/cold-chain-iot.json — it contains temperature logger data from 500 reefer truck trips across major US produce corridors: Salinas → Los Angeles (lettuce, strawberries), Yakima → Seattle (apples), Imperial Valley → Phoenix (winter vegetables), and Homestead → Atlanta (tomatoes). Temperature excursions, door open events, and GPS deviations are labeled. The modeling task: predict delivery quality score (Brix, firmness, color grade) from the IoT trajectory.

Cold chain anomaly detection uses two complementary approaches:

# 1. Rule-based CCP alerts (immediate, interpretable)
def check_temperature_excursion(readings: list[float], ccp_limit: float,
                                 duration_threshold_min: int = 30) -> dict:
    consecutive_violations = 0
    max_consecutive = 0
    for temp in readings:  # 5-min interval readings
        if temp > ccp_limit:
            consecutive_violations += 1
            max_consecutive = max(max_consecutive, consecutive_violations)
        else:
            consecutive_violations = 0
    excursion_duration_min = max_consecutive * 5
    return {
        "excursion": excursion_duration_min > duration_threshold_min,
        "max_excursion_min": excursion_duration_min,
        "estimated_shelf_life_reduction": excursion_duration_min * 0.8  # hours lost per min excursion
    }

# 2. ML-based predictive quality degradation
# XGBoost model trained on historical IoT + quality lab data
# Features: mean_temp, temp_std, max_excursion_min, n_door_opens,
#           total_transit_hours, g_force_events, origin_precool_temp
# Target: delivery_quality_score (0-100)

Perishable Demand Forecasting

The core challenge in fresh produce forecasting is the interaction between weather-driven supply shocks and elastic consumer demand. Romaine and onion prices at terminal markets can swing several-fold within a single season — a swing that destroys retailer margins and triggers shrink. The forecasting problem has two layers:

Supply forecasting (what will ship to the DC):

Demand forecasting (what will the retail channel absorb):

Prompt: "Using the attached USDA AMS shipment and terminal price data [ams-shipments-2023-24.csv]
for onion at Salinas, Yuma, and Idaho shipping points, and NDVI trend data from the relevant
growing district for the season, forecast: (1) weekly shipments for the next 8 weeks, (2) price
range (modal, low, high) with 80% confidence interval, (3) probability of a price spike above the
seasonal 90th percentile, (4) optimal procurement quantity for a retailer wanting to minimize
stockout probability while limiting cold storage inventory to 200 tons. Assume current cold store
utilization 65%, holding cost $0.03/lb/day, and stockout penalty equivalent to 3× margin loss."

Farm-to-Fork Traceability: FSMA 204 and Blockchain

FDA's FSMA Section 204 (the Food Traceability Rule) mandates lot-level traceability and Key Data Elements (KDEs) for foods on the Food Traceability List, with full compliance approaching. The technical infrastructure for farm-to-fork traceability is maturing rapidly, anchored on industry and government systems:

AI adds value at the data reconciliation layer: matching ASN/EDI lot IDs with cold store inward records (often manual and inconsistent), reconciling weight discrepancies between farm gate and DC arrival, and flagging lots where origin claims cannot be verified (fraud risk for PDO/PGI or USDA Organic products like Vidalia onions, Napa Valley wine grapes, or Parmigiano-Reggiano).

# Traceability record structure
lot_record = {
    "lot_id": "CA_SALINAS_20240315_0042",
    "commodity": "romaine",
    "variety": "Green Forest",
    "farm_origin": {"ranch": "Spreckels", "county": "Monterey", "state": "CA",
                     "grower_id": "PTI_GLN_0614141000012"},
    "harvest_date": "2024-03-12",
    "ams_transaction": {"shipping_point": "Salinas", "ship_date": "2024-03-14",
                          "fob_price": 18.50, "quantity_cases": 220, "grade": "US No.1"},
    "cold_store_entry": {"store_id": "DC_CA_0084", "entry_time": "2024-03-14T16:30:00",
                          "entry_temp": 2.1, "entry_quality_score": 88},
    "iot_telemetry_hash": "sha256:a4f9c2...",  # Immutable hash of sensor log
    "dispatch": {"destination": "DC_AZ_PHOENIX", "vehicle": "CA-7G45-21",
                  "dispatch_date": "2024-03-16"}
}

Terminal Market Price Prediction

Terminal market price prediction is a high-stakes forecasting problem: shippers, processors, and large retailers use it for procurement planning; growers use it to decide whether to ship immediately or store. The data infrastructure is now sufficient for reliable models:

Features that dominate produce price prediction (in order of importance from USDA AMS historical analysis):

Prompt: "I have USDA AMS terminal market data for tomato at the Los Angeles wholesale market for
the past 36 months [attached]. Build a price forecasting model that predicts: (1) next week's
modal price with 70/90% PI, (2) next month's average price, (3) the key drivers of price movement
for this week's forecast (SHAP-based explanation in plain terms for grower-cooperative members who
are not data scientists). Highlight any model uncertainty flags — periods where the model should
not be trusted due to out-of-distribution conditions like COVID-era disruptions or hurricanes."

Key Takeaways