For national postal operators, the address field is the single biggest determinant of delivery performance. When 20-40% of inbound parcel addresses are incomplete, malformed, or ambiguous — common even in developed markets once e-commerce volume crosses national post capacity — first-attempt delivery collapses. Shipsy’s Address Intelligence Service converts free-text, partially formed, and misspelled addresses into delivery-usable records at national scale.
The finding: most address quality is invented at the sender, not fixable at the postman
Postal operators typically inherit address data from two sources: e-commerce merchants (who accept whatever the buyer types) and cross-border inbound (where address quality varies by origin country). Both are upstream. The postal operator cannot enforce input quality at the source.
The leverage, then, is at the normalisation step. A national post that ingests raw addresses, cleans them, geocodes them, and enriches them before the sortation decision is in a completely different operational regime from one that tries to repair addresses at the delivery office — where the courier is already holding the parcel.
Why address quality is deteriorating, not improving
Three trends make this worse, not better.
E-commerce volume growing faster than national address databases. National postal address registries update on multi-year cycles. E-commerce is generating new apartment complexes, gated communities, mixed-use developments, and informal settlements faster than registries can index.
Cross-border inbound is address-quality-lossy. An address typed in one script, transliterated at export, re-interpreted at import customs, and handed to the national post arrives with typos, missing lines, or ambiguous locality fields.
Customer apathy and UX patterns. Buyers increasingly leave landmark references, informal locality names, and unit sub-addresses out of checkout forms. Default form patterns don’t enforce what postals actually need.
What Shipsy’s Address Intelligence Service does
The Address Intelligence Service is a pipeline, not a single model.
Ingestion and tokenisation. Free-text address strings are parsed into tokens — house/unit, street, locality, sub-locality, postal code, landmark, city, country — using a multilingual NER model trained on national postal data.
Normalisation against national registries. Each token is reconciled against the national postal database and local authoritative sources (municipal registries, gazetteer data). Misspellings are corrected; informal locality names are mapped to canonical ones.
Geocoding with delivery-point precision. Addresses are geocoded to the delivery point, not the street centroid. Where delivery-point geocoding is unavailable, the service falls back to sub-locality polygons with confidence scoring.
Enrichment with delivery heuristics. Addresses are tagged with known delivery constraints — gated community access codes, apartment doorbell codes where captured, courier-specific access notes, known-bad-address flags for addresses that have previously failed — built from Shipsy’s driver-app feedback loop.
Confidence scoring and dispatch routing. Every normalised address carries a confidence score. High-confidence addresses go straight to sortation. Low-confidence addresses route to a verification queue (via Clara, outbound SMS/IVR) before dispatch, avoiding wasted courier trips.
Postal address intelligence control points
| Failure mode | Address Intelligence mechanism | Delivery impact |
|---|---|---|
| Misspelled street or locality | Fuzzy match to registry + correction | Sortation accuracy up |
| Missing postal code | Reverse-lookup from locality + landmark | Hub routing enabled |
| Ambiguous apartment/unit | Flag + pre-dispatch verification via Clara | Fewer blind visits |
| Cross-border transliteration error | Multi-script NER + normalisation | Inbound parcels deliverable |
| Unknown informal settlement | Polygon fallback + driver-app feedback learning | Coverage of non-registered areas |
What postal ops leaders should do in the next 90 days
Measure your own address quality curve. Classify the last 30 days of inbound parcel addresses into high-confidence (straight-through), medium (normalisation corrected it), low (required human intervention), and dead (could not deliver). The shape of that curve is your operational ceiling for FADR.
Next, diagnose where the low-confidence volume is coming from. For most national posts it is a specific channel — one marketplace, one cross-border corridor, one informal-settlement postcode range. Address intelligence delivers disproportionate value where that concentration is highest.
Finally, link address normalisation to your driver app. When a driver encounters a bad address, the correction (or the “no such address” flag) has to feed back into the address intelligence loop. Without this feedback, the service stops improving after deployment.
For how this feeds FADR, see postal first-attempt delivery rate. For the postal vertical story, see how Shipsy fits postal operators. For the underlying product, see Shipsy Last Mile Delivery.