CASE STUDY / 01ADDRESS INTELLIGENCE

Rethinking how a courier
understands addresses.

The journey from an increasingly difficult routing algorithm to a layered, continuously improving address infrastructure for Bangladesh.

ROLETechnical Lead / Full-Stack Engineer
TIMELINE1.5 years of R&D
SCOPEDhaka & Chattogram
STACKGolang · BigQuery · Nuxt · Kubernetes
01 / THE INFLECTION POINT

Incremental improvement
was no longer enough.

Correctly identifying the Last Mile Hub from a merchant's raw address is fundamental to courier operations. The original APT improved over time, but as parcel volume grew, further accuracy gains demanded disproportionate effort and misroutes remained costly.

Hands-on work with millions of real addresses revealed the deeper issue: people do not describe locations in a clean three-level structure. They use buildings, markets, roads, blocks, landmarks, local spellings, mixed scripts, and relationships that can extend to any depth. The solution needed a new mental model—not another rule.

02 / THE NEW MODEL

A map with no fixed
depth limit.

Every recognized place can connect to a parent, allowing the parser to travel from broad administrative context to the smallest useful delivery detail.

01BangladeshCountry
02DhakaDivision
03PanthapathArea
04Bashundhara CityBuilding
05Level 4Floor
06Block DBlock
07Shop 56Unit
03 / ONE SYSTEM, THREE PARTS

Infrastructure, operations,
and the parser.

01

IbnBattuta

A web application for discovering, visualizing, linking, and validating layered geographic areas from historical delivery data.

  • Place-name and alternative detection
  • Parent–child relationship suggestions
  • Verification task management
  • Noise-resistant historical benchmarking
02

Map Operations

A dedicated team validates suggested locations and their parents, turning algorithmic discovery into trustworthy geographic knowledge.

  • Human verification loop
  • Non-technical correction workflow
  • Fast benchmark-and-fix feedback
03

APT 2.0

A next-generation parser for informal, multilingual, misspelled, ambiguous, and incorrectly ordered address input.

  • Transliteration and spell correction
  • Custom inverted indexes
  • Configurable rules engine
  • Validity and hierarchy checks
04 / THE HARD PART

Bangladesh does not
fit a dropdown.

Addresses are informal, multilingual, landmark-heavy, misspelled, and sometimes contain places in the wrong order. The same road name can exist in many areas, while a familiar area word can be misleading.

Hierarchy ambiguityCommon words can refer to unrelated places.
Local knowledgeLandmarks often carry more signal than formal labels.
Changing operationsHub boundaries must evolve without algorithm rewrites.
Layered
place graph
70M+Early address records
MineDiscover places
VerifyHuman operations
BenchmarkHistorical truth
ConfigureImprove safely
05 / EDGE CASES

What the parser
learned to understand.

01
RAW INPUT

Pathao Hedkuatar, Gushan Dhaka Gushan Lake er pashe

RESOLVED HIERARCHY
DhakaGulshan 2Kamal Ataturk AvenuePathao HQ
Mixed language, misspelling and landmark context
02
RAW INPUT

Bashundhara Block D L4 Panthapath

RESOLVED HIERARCHY
PanthapathBashundhara CityLevel 4Block D
Disambiguates a landmark from a similarly named area
03
RAW INPUT

Shahbag Thana

RESOLVED HIERARCHY
ShahbagDhaka UniversityShahbag Police Station
Understands that “Thana” describes a specific place
04
RAW INPUT

Fazazy Hospital er pichone Gulshan Badda Notun Bazar

RESOLVED HIERARCHY
DhakaBaddaMadani AvenueFarazy Hospital
Uses a misspelled landmark to resolve conflicting locality words
06 / FIRST PRODUCTION DEPLOYMENT

Accuracy that changed
operations.

98.8%Hub detection accuracyCorrect routing in the first Dhaka deployment
99.7%Address coverage997 of every 1,000 addresses detected
99.80%Nearby-hub accuracyOnly one major routing error per ~500 parcels
14msCurrent mean latencyImproved from approximately 20ms during the earlier rollout

After deployment in Dhaka and Chattogram, hub exchanges dropped dramatically. Configuration-driven behavior also made it possible to adjust boundaries, introduce hubs, and handle new patterns without repeatedly changing the core parser.

~2M BDToperational savings / month
THE OUTCOME

Not an algorithm upgrade.
A foundational shift.

A complete feedback loop now collects addresses, discovers place relationships, supports human curation, benchmarks changes, and feeds better intelligence back into parsing. As deliveries grow, the map learns new areas, patterns, and behavior.

Return to selected systems