DataTex: In-House Address Intelligence Engine

Fix the root cause of a large share of delivery failures: poor address data. Addresses arrived from many different shippers in inconsistent formats, with no method to interpret, standardize, consolidate or correct them. This caused 7% delivery non-conformity, re-delivery and return costs, 14% more customer service tickets and a loss in on-time delivery (SLA). Goals: (1) build an in-house engine that validates, standardizes, enriches and geocodes every address before routing; (2) raise on-time delivery and cut customer complaints; (3) make routing more accurate and reduce operating costs; (4) replace expensive third-party address tools with a low-cost, serverless solution; and (5) design it so it can later be offered to the market as an address-treatment service.

Results. On-time delivery (SLA) up from 95% to 96%; customer complaints down 5%; more accurate geocoding and routing, fewer re-deliveries and lower operating costs; third-party tooling cost cut from R$ 40,000 to about R$ 100 a month (-99.7%).

• DataTex intelligence engine – hundreds of in-house functions for validation, standardization and consolidation

• Amazon Bedrock (LLM) – exception handling, address classification, attribute splitting and house-number extraction

• AWS Lambda – serverless processing functions

• Amazon SQS – event-driven information flow

• Amazon DynamoDB – storage of data and quality metrics

• Big Data / Analytics – address history analysis and pattern identification

• Proprietary enrichment bases and geolocation – latitude/longitude for treated addresses

• Systems integration – connection with internal company systems and external data sources

Brazilian addresses are notoriously irregular, and the engine had to handle data from many different shippers, each with its own formats and errors, without slowing the operation. Key difficulties: (1) turning tacit operational knowledge into hundreds of reusable, configurable functions; (2) integrating heterogeneous internal and external sources into one consistent record; (3) deciding when deterministic rules are enough and when to escalate to an LLM, keeping cost and latency under control; (4) building proprietary enrichment bases from historical delivery data; (5) improving geocoding precision; and (6) running at full operational volume on a serverless architecture costing about R$ 100 a month. Requirements were refined iteratively as the team learned from real data, with continuous measurement to keep data quality high.

Context. Total Express is one of Brazil's largest logistics and e-commerce delivery companies. Every parcel starts with an address sent by a shipper, and in Brazil these arrive in highly inconsistent formats: missing numbers, abbreviations, swapped fields, free text.

Challenge. With no method to interpret, standardize, consolidate and correct address data, errors flowed straight into routing and delivery. The result was 7% delivery non-conformity, re-delivery and return costs of about R$ 10,000 14% more customer service tickets and a loss in on-time delivery. The company also paid around R$ 40,000 a month for third-party address tools.

Solution. IT built DataTex, an in-house address intelligence engine that treats every address before routing. It combines internal systems with external sources and adds new data such as geolocation.

1. Rules engine. A library of hundreds of functions validates, standardizes and consolidates addresses, driven by business rules configurable per customer.

2. Proprietary enrichment. Big Data analysis of the company's own address history identifies patterns and builds enrichment bases that improve classification.

3. GenAI for exceptions. A large language model handles only the cases rules cannot resolve: classifying the address, finding the house number and splitting free text into structured attributes.

4. Geocoding. Treated addresses receive more precise latitude and longitude, improving routing.

5. Continuous improvement. Quality metrics are stored and reviewed continuously, and rules and models are adjusted based on results.

Architecture. Fully serverless and event-driven on AWS: Lambda functions process addresses, SQS manages the information flow, DynamoDB stores data and metrics, and Amazon Bedrock provides the LLM. Cost scales with volume.

Method. Process and gap mapping, agile delivery (Scrum/Kanban), proofs of concept for data treatment and geolocation enrichment, integration tests with heterogeneous sources, team training. No external consultancy was used.

Results. On-time delivery (SLA) up from 95% to 96%; customer complaints down 5%; more accurate geocoding and routing, fewer re-deliveries and lower operating costs; third-party tooling cost cut from R$ 40,000 to about R$ 100 a month (-99.7%).

Next steps. Productize DataTex as an address-treatment service for the market, with only minor adjustments.

Brazil (nationwide delivery operation)