Shipment status, exceptions and proof of delivery.

Start with the enquiry types and the systems that answer them

List the requests that actually arrive: where is my consignment, why was it not delivered, move the delivery to tomorrow, send me the proof of delivery. Order management holds the order, the transport management system the consignment and trip, the warehouse system dispatch state, the carrier most of what happens after handover.

One order becomes several consignments that can cross a line-haul carrier and a last-mile partner, so authority between disagreeing systems is settled first.

Status data arrives late and in inconsistent shapes

Status is a stream of scan events, not a single field: picked up, inbound to hub, out for delivery, undelivered, delivered. One carrier sends webhooks, one a rate-limited polling API, one an overnight SFTP file, one EDI status messages, and the same event appears under different codes at different carriers.

Events arrive out of order and twice over when a handheld syncs late, and event time is unreliable: clocks drift, hubs stamp local time with no zone, batch uploads backdate. So codes map into a canonical milestone model, deduplication uses a composite key over consignment, code, location and timestamp where no stable event id exists, ordering takes event time with received time as tie-break and milestone precedence where a timestamp is unusable, and unmapped codes surface as unknown.

Resolve the reference, check who is asking, then answer within the feed

A shipment carries several identifiers: a tracking number, an air waybill, a consignment note or LR number, the shipper order reference, sometimes only the caller's number, and a reference can match several consignments or one piece of a multi-piece shipment.

Who may ask is settled before what may be answered. Calling line identity narrows the search but does not authorise it, so a read is released only against a detail the caller supplies and the record holds. An address change, a hold or a redirect needs a stronger check, because diverting a parcel is the fraud these flows attract; authority comes from the carrier contract and is logged with the write.

The answer is bounded by the feed: the last confirmed scan and when it was recorded. A next update time exists only where a carrier drops scheduled files. A delivery time belongs to the carrier or the route plan, restated with its source and never as the assistant's own estimate.

Exceptions, reattempts and rescheduling a window

Exception traffic is where the handling effort concentrates, and it is not one case. Consignee unavailable, cash not collected on a cash-on-delivery consignment, damaged in transit, held at a checkpoint for documentation and missorted to the wrong hub each have a different owner. Each code carries its permitted actions: reattempt within contract limits, address correction inside the carrier edit window, or route to the control tower.

Classifying and routing these is in scope; deciding them is not. Damage liability, shortage settlement, release of a shipment held for documentation and movement of cash-on-delivery money stay with their owners; the case is routed with its evidence.

Capacity for a reschedule sits in route planning or a carrier slot API with its own cut-offs, and once a trip is dispatched the change must reach the driver, not the plan. Write actions carry a per-request idempotency key over the consignment reference, the action and the slot, so a retry does not duplicate a redelivery order while a genuinely different change still goes through.

Callers in vehicles, warehouses and patchy coverage

Driver and field-staff calls arrive in hard audio conditions: a cab at speed on speakerphone, a loading bay with forklifts. A consignment number, vehicle registration or seal number fails on one wrong character, selecting a different shipment rather than failing visibly.

So prompts stay short, digits are read back before any write, and expected values are narrowed to the stops and piece labels on that driver trip. A number that cannot belong to that fleet is rejected rather than matched to the nearest, and call state survives a dropped connection so the caller resumes rather than restarts.

Proof of delivery and the documents a dispute is argued from

Proof of delivery is evidence a billing or shortage dispute is argued from, alongside the piece count accepted and remarks written at handover, so capture is held to a documentary standard. The form varies: a signature on a handheld, a photograph at the door, a stamped delivery challan, or a consignment note photographed days later by a subcontracted fleet.

The document is matched to the right consignment and piece count, and the signatory checked against the consignee or an authorised recipient the carrier contract allows, the relationship recorded rather than treated as a mismatch, because a neighbour, a guard or a mailroom signing is ordinary. Anything missing is flagged with the image.

Integration constraints and what happens when a feed is down

Alongside REST APIs from carriers and aggregators there are EDI transaction sets for load tenders and status messages, EDIFACT on international lanes, and flat files over SFTP. A warehouse system may be on-premises, reachable only through a database view or nightly extract.

Polling against a rate limit means the status served is cached, and the cache age belongs in the answer. Whether test environments carry representative consignments is checked rather than assumed, and the sprint settles what runs live and what is replayed.

Degradation is defined per dependency. With a carrier API unreachable, the system answers from the last stored event and says how old it is. With the transport system unavailable, a reattempt is queued and the caller told it is requested rather than booked, while a window change against a passed cut-off is refused rather than queued to drain late.

Evaluate against real consignment histories

The evaluation set comes from real tracking histories: no scan for several days, events out of order, a piece reassigned to another last-mile partner mid-route. Identity resolution, milestone correctness, exception classification and refusal to answer are scored separately, with driver calls on vehicle and warehouse recordings.

What a logistics project costs

Every engagement starts with a fixed-price discovery sprint at USD 2,500 over two weeks, which produces the architecture, initial screen designs, acceptance criteria, a running-cost model and a fixed price for the build. The sprint fee is deducted from the build.

Running cost has three parts that do not move together: contact handling per minute for voice and per exchange for messaging, document processing per proof of delivery page tracking delivery volume, and status ingestion, which runs whether anyone asks because carrier events arrive continuously.

Part of the AI Feature Implementation Sprint → · Explore our work → · Discuss an AI project →