Vendemmia is a 4PL logistics operator that integrates and manages the entire supply chain for companies importing goods from abroad. Its portfolio covers importation, warehousing, transportation and logistics management, coordinating the flow of international cargo from pickup at origin through customs clearance and final delivery.

Challenge

Onboarding that was manual, slow and without a standard

Every import process generates dozens of documents, for example: tax invoices, commercial invoices, packing lists, articles of incorporation, company tax ID records, tax clearance certificates, electronic powers of attorney, import and export licenses, cargo spreadsheets, and vessel and customs clearance information.

Before the project, registering a new client required uploading every piece of information manually, document by document and field by field, into the company’s system. On top of that, each client followed its own standard, when there was one at all. Documents arrived in different formats (PDF, JPEG, XLSX) and languages (Mandarin, French, English and Portuguese), creating three operational challenges:

• Constant rework to standardize information that arrived with no standard;
• An operational bottleneck that limited how much the client base could grow;
• Failed attempts to solve it with traditional database automation, which could not handle the enormous variation between documents.

It was in this scenario that Programmers stepped in to help Vendemmia turn its data operation into a modern, scalable, AI-supported structure.

From proof of concept to AI in production

The starting point was a proof of concept applying a medallion architecture and native ETL processing inside the Databricks Lakehouse. Next, to enable direct and simple access to the data, we enabled Databricks Genie, a tool that allows natural language queries with visual answers.

It was during a conversation with Genie that the central idea of the project was born. After all, would it be possible to use AI not only to query already structured data, but to extract information directly from clients’ raw documents?

The first attempt was to integrate external AI models, such as GPT and Claude, through Databricks endpoints. It worked, but it required hundreds of lines of code that were complex to maintain. The turning point came when Databricks released Agent Bricks, a tool that makes it possible to create specialized AI agents without programming from scratch. In other words, what used to take hundreds of lines of code became a single configuration.

Vendemmia began the project thinking about using AI only to visualize data. It ended with AI agents in production, processing thousands of documents per day.

Solution

How the platform works

The architecture runs on Azure, combining ADLS as the raw layer, Delta Lake for structured and transactional data, and a hybrid execution model, event driven for triggers and batch for recurring processing. All processing and governance run on Databricks (Spark, Delta Lake, Lakeflow, Agent Bricks and Unity Catalog).

The pipeline built by Programmers works in three main stages:

1. Reading the documents

Using the AI Parse Document function from Agent Bricks, the system automatically extracts text from images and PDFs in any language, with native detection and translation, without a single line of custom code. Each document becomes a row in a table, with a column identifying the source language.

2. Extracting the right information

A custom AI agent receives a list of around 15 target fields, such as the client’s company tax ID, the destination warehouse tax ID, container number, invoice number and bill of lading number, among others, and returns everything already structured. The agent went through careful fine tuning: early on, for instance, it confused invoice numbers with container numbers, but by refining the description of each field it learned to tell them apart correctly.

3. Consolidation and business rules

Because a single import process can involve several emails exchanged over days or weeks, the system recognizes that those messages belong to the same process and consolidates everything into a single record. Vendemmia’s business rules are then applied to that record, marking each piece of information as final, pending or provisional, and updating the record automatically as new emails arrive.

Because the process involves personal and sensitive data, the entire privacy layer was built with Unity Catalog, the Databricks data governance module, with masking of sensitive information, fine grained access control and compliance with data protection regulations. Any consolidated piece of information can be traced back to the document and the exact excerpt it came from, turning audits that would take days into queries answered in seconds.

Results

From months to days, from days to seconds

Onboarding up to 6x faster. What used to take 1 to 3 months per client now takes between 1 day and 2 weeks, depending on the complexity of each client’s business rules.

Processing scale. The first full data load for the first client served processed around 1,500 documents. The target is for the platform to reach between 10,000 and 15,000 documents per day.

A commercial differentiator. Automating process registration was a decisive factor in helping Vendemmia win new clients. Instead of typing information field by field, the end client now only needs to scan the documents and send them by email.

Audits in seconds. With full traceability back to the origin of every data point, audits that used to take days are now resolved in seconds.

“For our client, the change was easy to feel: instead of typing information field by field, they simply scan the documents and send them by email. For us, it was a structural change, from constant rework to reliable, traceable data. Today we can grow our client base without stalling the operation, and that was decisive, including in winning new clients.”Luciano Ricci, Founder Director at Vendemmia

What comes next

The platform is in production today with the first client served, but Vendemmia’s ambition is bigger: taking the same pipeline to every client and to other areas of the company, such as finance, creating agents for each type of document.

The same foundation, governed, reusable and already validated in production, enables a series of other solutions: expense reimbursement systems with automatic reading of receipts, personal document management for compliance and KYC, automation of operational processes, and alert and event systems. Following this project, Vendemmia decided to build its entire analytics platform on Databricks, moving away from treating data and AI as a one off project and making them a permanent part of the engine of the business.

“We built this solution to solve onboarding for one specific Vendemmia client, but what came out of it was a reusable foundation. That is the kind of architecture that turns into real competitive advantage.” Railson Lima, Customer Success Manager at Programmers

Want to know how Programmers can help your company turn unstructured data into competitive advantage with Databricks and AI? Get in touch now.