Systems that model the physical world, where recorded state diverges from reality by default and reconciliation is continuous rather than exceptional.
18 published across 7 topics.
The end-to-end flow, milestones, and the event model that reports it.
It depends entirely on where it stopped, because silence is normal on a long line haul and alarming between two hand-offs that should be minutes apart. The design that answers it is an expected next event with a deadline per milestone, rather than a threshold on time since the last scan.
Model the journey as milestone events, each carrying both the time it happened and the time you recorded it, and derive current status by projecting that set in a fixed milestone order rather than storing a mutable column, because events arrive late, twice and in the wrong sequence.
The split forces order, shipment and parcel apart as separate entities with their own promises, money and tracking, so a single order status stops being expressible. Model the promise per line, capture funds per despatch, and give the customer a page that reads as one order made of several deliveries.
Receiving, putaway, picking strategies, packing and the WMS.
Warehouse picking strategy depends on order profile, SKU velocity, layout and travel time. A WMS becomes real-time when picker task dispatch, stock exceptions and replenishment decisions must respond while people are waiting in the aisle.
Block the location first so nothing allocates against a figure you have just disproved, then correct it as a recorded adjustment with a reason rather than an edit. Accuracy is maintained continuously by cycle counting, because stock records drift even in a well-run warehouse. It also connects inventory accuracy to the point an interviewer is testing.
Load planning, vehicle routing as a heuristic problem, and carrier selection.
Triage before optimisation: classify what is already committed, what can still be diverted and what has not shipped, then reallocate the scarce alternative capacity by business priority rather than by cost. It only works if your model holds legs and alternates rather than a single route with a port baked into it.
Usually nothing is wrong with the solver: it is optimising against a model missing constraints the planner knows and never told it, so the plan is optimal and unrunnable. The fix is to capture the overrides as data and turn the recurring ones into constraints, rather than to improve the objective.
You stop pursuing optimality and engineer for a wall-clock budget: build a feasible plan with a construction heuristic, improve it with local search or a metaheuristic until the clock runs out, and model the constraints that genuinely bind, which are time windows, capacity and statutory driver hours.
Slotting and delivery promises, driver apps, proof of delivery and failed attempts.
A slot shown at checkout has to be a reservation against remaining route capacity measured in minutes rather than a count of orders, and the flow behind it needs an offline-tolerant driver app, proof of delivery sized to the risk, and a designed path for the attempt that fails.
Treat the device as a first-class writer rather than a cache: it generates its own identifiers, records two timestamps per event, and queues durably until the network returns. Every submission carries an idempotency key because retries are certain, and genuine contradictions go to a human.
A failed delivery attempt in last-mile logistics is a structured event, not just failed status. The system must capture the allowed outcome, proof, attempt budget, customer notification and next custody location so redelivery, pickup or return can proceed safely.
Scan events, estimated arrival, exception detection and customer communication.
Translate scans into a small set of customer-meaningful states, expect long silences and out-of-order uploads, never let a displayed status move backwards, and publish an arrival estimate with its uncertainty made visible rather than a single confident time you will miss.
Rebuild rather than reconstruct: most of those events still exist in carrier systems, EDI mailboxes and handheld queues, so re-ingest them idempotently by date range, then close the residual gap with physical observation at yards and terminals. It only works if raw inbound feeds were archived apart from the projections built from them.
Demand forecasting, safety stock, replenishment and the bullwhip effect.
Accuracy is almost always reported at a grain and horizon that inventory does not care about, so a national monthly number can improve while error at item and location over the replenishment lead time gets worse. Bias, weighting and the history you trained on explain the rest.
Safety stock is the inventory you hold to buy a chosen service level against forecast error over the replenishment lead time, so it scales with the variability of demand and of lead time; the bullwhip arises because each tier reorders from its own buffered demand rather than the real one.
Cross-border documentation, duties, and compliance screening.
Lodged is one state in a sequence and the broker is reporting the only one they volunteer, so you have to enumerate the states, demand a timestamp and a reference for each, and age them. The reason it matters is that storage and container charges accrue from the moment the box lands, and nobody in the chain owns that clock.
The code sets the duty rate, the controls that apply and whether the goods can move at all, so two codes mean two different amounts of tax paid on identical goods. It is the importer's legal responsibility rather than the freight forwarder's, and inconsistency is what an audit looks for.
The goods need a commercial invoice and packing list that agree, a defensible tariff classification and country of origin, an Incoterm fixing who pays duty and tax, and a completed restricted-party and export-control screen that blocks release rather than warning after despatch. Use this customs answer to show the decision, trade-off, and evidence rather than a memorised definition.