With support from AWS, Bravante RRC uses generative AI to find deactivated oil pipelines

The process is regulated by the National Petroleum Agency, in resolution 817/2020, which states that the seabed must be left as it was found after exploration. Felipe Santello, product manager at RRC, explains that since there are pipelines that have been there for over 30 years, finding them depends on analyzing documents that are often scattered and in different formats. It is a task carried out by naval specialists who analyze them and determine where this equipment should be found. The issue is that the locations are not always accurate and the daily rate of the ship carrying out the certification is US$ 250k, which implies a loss if the equipment is not actually found and removed.

“The purpose of the vessel is to verify that the pipeline is in its documented state, verifying the weight, condition of the stretch and its existing marine life, which also needs to be preserved,” he explains. “To give you an idea, each stretch has six or seven reports - 15 to 40 pages each - in addition to 2D and 3D images that need to be read and analyzed. Considering the more than 800 km of pipelines to be decommissioned, the data repository has over 10 thousand documents.”

To reduce the manual work involved in the analysis of such a complex repository and increase the accuracy of the locations, the company decided to develop a solution based on Generative Artificial Intelligence (AI) to analyze documents automatically and determine the exact location of the parts that must be removed. “Decommissioning has been increasingly required by the ANP and there is a need to gather all information in a single trustworthy source. We saw the existing potential.”

Amazon Web Services, which had been investing in the relationship with Grupo Bravante for months, was interested in the challenge and offered the company its support to develop and test an artificial intelligence (AI) platform. Flexa Cloud, an official AWS partner with experience in corporate generative AI projects, was invited to build this solution with Bravante RRC scientists.

The architecture built starts with the creation of the document repository in Amazon S3, allowing fast, secure, and low-cost access to thousands of documents. With native S3 object integration, Amazon Textract was used for optical character reading (OCR). The versatility of the service was crucial considering that the pipeline document repository includes images, spreadsheets, pdfs, and other formats.

“After extracting this data, we used the semantic capacity of a foundational model to interpret and classify the data at scale,” explains Flexa Cloud CEO, Deivid Bitti. “We started off with models that were customized using Amazon SageMaker, which brought us a lot of computing power. Once we had access, we switched over to Amazon Bedrock architecture.”

In the view of Deivid Bitti, “Bedrock brings great versatility in the choice of foundational models. It facilitated the development and experimentation process by being quickly available via APIs.” Felipe Santello adds that “by using advanced interaction techniques with the foundational model, we achieve high precision without the need for prior training”.

The foundational model chosen was Anthropic's Claude-2. “The context window was a very important differential. We need to keep the model's attention for each stretch, which involves up to 7 reports with up to 40 pages each. The 100k token window was critical to our objective,” says Felipe.

Data extraction and processing via Amazon Bedrock allowed for the creation of a complete dossier on each of the stretches in an Amazon Aurora-structured database (DB). Bitti reveals that, with the data provided by RRC, Flexa developed an preliminary platform from which Bravante could use and validate the solution. As a next step, “we are testing a technology that can describe images in documents, in a scalable way, and enrich the data available to Bravante's naval engineers,” concludes Deivid Bitti.