Tom monitors every freight lane, every vessel, every weather system, and every geopolitical risk — simultaneously, continuously. It sees disruptions 72 hours before they happen and reroutes autonomously in 14 minutes. No emergency calls. No chaos. Just delivery.
Tom doesn't react to disruptions. It predicts them, routes around them, and keeps Becci's production schedule intact.
Tom fuses satellite, meteorological, geopolitical, and commercial data into a single real-time risk picture across every lane.
A pre-recorded simulation of Tom detecting Typhoon Kai approaching the Taiwan Strait, calculating risk to 14 active shipments, and executing a full reroute through Singapore — all in 14 minutes, at 2am, without a single human decision. This is what your supply chain looks like.
Configure your freight operation below and every response becomes specific to your lanes, carriers, and disruption scenarios.
This paper presents Tom, the autonomous logistics agent within the VYUH Supply Chain Neural Network. Tom employs an ensemble disruption prediction model — combining gradient-boosted classification, LSTM sequence modelling, and Bayesian geopolitical risk propagation — to forecast freight lane disruptions an average of 72 hours before they become critical. Upon disruption detection, Tom solves the rerouting problem as a Dynamic Vehicle Routing Problem (DVRP) and executes the optimal reroute autonomously within 14 minutes. We describe the full technical architecture, prediction model design, DVRP formulation, confidence gating mechanism, and integration within the VYUH agent network.
Global freight networks are inherently exposed to disruption — typhoons, port strikes, geopolitical restrictions, vessel failures, and carrier capacity constraints cause regular service interruptions across every major trade lane. The economic cost of these disruptions falls disproportionately on companies that react to disruptions rather than anticipating them. Emergency freight premiums, expediting costs, and production downtime caused by delayed materials represent a significant and largely avoidable cost category.
The fundamental problem is one of information and response speed. Disruptions are often predictable days in advance from publicly available data sources — weather models, port operations feeds, AIS vessel tracking — but the integration and interpretation of these signals across hundreds of active freight lanes exceeds human cognitive bandwidth. By the time a logistics planner identifies a developing disruption and begins evaluating rerouting options, the low-cost rerouting window has often closed.
Tom addresses this through continuous, automated monitoring and instant autonomous response — detecting disruptions from multi-source signals, solving the rerouting problem mathematically, and executing the optimal solution before the disruption window closes.
Tom's architecture comprises four layers: (1) a multi-source data ingestion layer processing AIS vessel feeds, NOAA/ECMWF weather models, port authority data, carrier APIs, and geopolitical risk feeds in real time; (2) an ensemble disruption prediction model that synthesises these inputs into per-lane risk scores with 72-hour forecasting horizon; (3) a Dynamic Vehicle Routing Problem (DVRP) solver that calculates optimal rerouting solutions when disruption is confirmed; (4) a confidence-gated execution layer that books carriers and notifies inter-agent peers autonomously for high-confidence solutions.
Tom's disruption prediction model is an ensemble of three components. The primary component is a gradient-boosted classifier (XGBoost) trained on a 5-year historical dataset of lane disruptions paired with pre-disruption signal features — including weather model outputs, port congestion metrics, AIS anomaly patterns, and geopolitical risk indices. This model achieves 89% precision and 84% recall at the 72-hour horizon for major disruptions.
The second component is an LSTM sequence model that processes time-series signals — particularly AIS vessel behaviour patterns and port throughput trends — to identify temporal anomalies that precede lane disruptions but are not captured by static feature models. The LSTM adds approximately 8 percentage points of recall for disruptions with gradual onset (port congestion, slow-developing weather systems).
The third component is a Bayesian network for geopolitical risk propagation — modelling how political events in one region propagate through connected shipping routes and port operations. This is particularly relevant for trade restriction events and port industrial action, which spread through the network in predictable patterns.
The three model outputs are combined through a learned ensemble weighting that is recalibrated quarterly against recent prediction performance. The final disruption probability estimate triggers the DVRP solver when it exceeds a lane-specific threshold calibrated to the cost of unnecessary rerouting versus the cost of a missed disruption.
When disruption probability exceeds threshold, Tom formulates the rerouting problem as a DVRP. The problem is defined over a graph G = (N, E) where nodes represent ports and intermediate waypoints, edges represent freight lanes with associated cost, time, and capacity attributes. The objective is to find the minimum cost routing for all affected shipments that satisfies delivery time constraints and carrier capacity constraints.
The DVRP is solved using a two-phase approach: an initial solution is generated using the Clarke-Wright savings algorithm, then improved using a tabu search metaheuristic with a 10-minute computation budget. This consistently produces near-optimal solutions — within 3% of the mathematical optimum — while meeting the 14-minute total response time target (which includes data ingestion, problem formulation, solving, and carrier booking).
Rerouting solutions are subject to a composite confidence gate before autonomous execution. The confidence score integrates: disruption model certainty (how confident is the ensemble model in the disruption prediction?), DVRP solution quality (how close is the solution to the theoretical optimum?), carrier booking confirmation (has the alternative carrier confirmed slot availability?), and production schedule impact assessment (has Becci confirmed the revised timeline is acceptable?). Solutions above 82% confidence execute autonomously. Between 65% and 82%, a logistics team review flag is raised with a 30-minute window. Below 65%, a full escalation briefing is generated.
Tom's integration with the broader VYUH agent network is essential to its production-protecting function. Becci's production schedule identifies the critical path materials — the specific inbound shipments whose delay would cause a production halt. Tom applies heightened monitoring to these shipments and lower disruption thresholds, ensuring that critical materials receive the earliest possible warning and the most aggressive response. When a reroute modifies a critical-path delivery timeline, Tom notifies Becci immediately with a revised ETA — triggering proactive replanning rather than an emergency response. Cho receives revised inventory arrival timing, updating the stockout risk model accordingly.
Over 12 months of operation across a 340-lane global freight network: average disruption warning time 72 hours (vs industry standard of same-day reactive response). Average reroute response time 14 minutes. Emergency freight cost savings: £6M per major disruption event. Freight cost reduction through continuous route optimisation: 23%. Carrier SLA degradation caught pre-failure: 94% of cases (vs 31% in manual monitoring environment). Production delays caused by logistics failures: reduced by 87%.
Tom's disruption prediction model has lower accuracy for novel event types with no historical precedent in the training data. The 72-hour warning horizon reflects average performance — some disruption types (sudden geopolitical events, unexpected port industrial action) occur with shorter warning windows, and Tom's response in these cases is reactive rather than anticipatory. The DVRP solver assumes carrier slot availability can be confirmed rapidly; in periods of extreme industry-wide disruption, carrier capacity constraints may extend the booking confirmation step beyond the 14-minute target. These limitations are disclosed transparently in every disruption response briefing generated by Tom.