Address Intelligence: how Shipsy turns unstructured addresses into deliverable jobs
A bad address is the most expensive thing in a parcel network. Shipsy’s Address Intelligence Service takes a free-text, unverified address and returns a geocoded, deliverable shipment — materially lifting address-hit-rate on messy corpora. For postal operators, that is the difference between a profitable route and a route that needs a second attempt.
This deep-dive is for postal and CEP operators whose geocoding stack already works for “clean” addresses and fails at exactly the 15–30% of addresses that actually cost them money.
Why we built this
Most geocoders are trained on clean, Western, single-line addresses. A real postal network handles apartment numbers written in three scripts, landmarks instead of street numbers, 7-digit postcodes where only the first 4 are valid, and shipper-provided addresses that look like “Near the mosque, opposite Daily Bread, second gate.” Off-the-shelf services fail silently on these and return low-confidence matches that look fine in a dashboard but fail in the van.
Shipsy’s customers in postal, CEP, and quick commerce were bleeding a meaningful share of last-mile cost on failed deliveries traceable to address quality. We built Address Intelligence as a pipeline, not a single API — because fixing an address is a multi-stage problem.
How it works
The Address Intelligence Service (AIS) runs as a deterministic pipeline with four stages. Each stage annotates the shipment and passes it forward. Human-in-the-loop review is triggered only when confidence falls below a tunable threshold.
1. Parse and standardize
The first stage decomposes free text into structured fields — unit, building, sub-locality, locality, postcode, city, country — using locale-specific parsers. Transliteration, diacritics, and abbreviation expansion happen here. The parser is country-tuned because “Block 7A” means very different things in Singapore, Delhi, and Warsaw.
2. Verify and enrich
Parsed components are validated against authoritative postal databases where available, against Shipsy’s proprietary delivery-history index where not. If a shipment arrives for “B-42, Gulshan,” Shipsy has likely delivered to that building many times and knows its canonical geocode, the gate to use, and the typical delivery window.
3. Geocode with confidence scoring
The address is geocoded to rooftop precision where possible, street segment otherwise, and every geocode ships with a confidence score and a provenance tag — did this come from a public geocoder, from Shipsy’s history, or from operator confirmation?
4. Route and close the loop
Addresses below a confidence threshold are surfaced to operators for correction before dispatch. Every successful delivery feeds back into Shipsy’s history index. Addresses that fail are flagged, and the correction loop improves the parser for the next shipment in that locality.
The net effect: addresses improve over time inside the network, without the shipper ever touching their upstream system.
The pipeline looks like this:
Early results
Postal customers running Shipsy AIS see address hit-rate climb materially on legacy corpora during the first operating quarter, with the biggest gains concentrated in exactly the dense urban and informal-addressing geographies that had been bleeding cost. Qatar Post now runs 90% first-attempt delivery rate with 12–18% cost reduction on Shipsy; IKEA hits 95% FADR on big & bulky. Quick-commerce operators see a different benefit entirely: AIS pre-resolves ambiguous addresses at checkout, shaving precious seconds off the promise-to-dispatch window.
For enterprise shippers running Shipsy’s TMS alongside AIS, the value shows up as higher first-attempt delivery rates and lower customer-service ticket volume on “wrong address” disputes.
What’s next
The next phase is tighter integration with Clara, our CX agent, so low-confidence addresses can trigger a proactive WhatsApp or SMS clarification to the recipient before the shipment leaves the depot — turning a failed delivery into a successful one before the driver ever gets in the van.