How Azati turned millions of complex engineering documents into searchable and structured data for a global energy company
A Fortune 500 energy company operating across more than 50 countries needed to turn a large technical-record estate into searchable and structured information. The estate included engineering drawings, regulations, contracts, and pipeline and facility documentation.
Existing processing relied heavily on manual handling and did not deliver the required extraction quality. Azati built a distributed document intelligence pipeline that processed documents from raw PDFs through OCR, classification, metadata and entity extraction, validation, QA/QC, and structured reporting.
The goal was not simply to make PDFs searchable. It was to create a repeatable processing system that could produce controlled, operationally usable data at industrial scale.
technical documents in scope
distributed processing jobs
pipeline processing steps
Case study at a glance: Azati engineered a distributed document intelligence pipeline for a global Fortune 500 energy company, processing approximately 5 million technical records across 50+ countries.
The system combined custom OCR, document classification, domain-specific entity extraction, QA/QC, reconciliation, and structured reporting.
The business challenge: turn a large, heterogeneous technical-document estate into searchable, structured information without relying on manual processing.
The solution had to handle approximately 5 million files and terabytes of data, requiring distributed processing rather than conventional single-machine OCR.
The required output went beyond plain text. The pipeline needed to identify document types, titles, Authorization for Expenditure (AFE) and Facility Tracking Numbers (FTNs) information, engineering metadata, and named entities relevant to pipeline and facility records.
Piping and instrumentation diagrams (P&IDs) and technical drawings can contain rotated content, polygon-shaped text regions, high-resolution images, and unusually long text lines that challenge standard OCR approaches.
The client needed more than a model prediction. Business rules, confidence scoring, out-of-distribution detection, reconciliation, and QA/QC outputs were incorporated into the processing workflow.
The differentiator wasn't another OCR model. It was the engineering around it.
For millions of technical records, OCR alone could not provide the structured information the client needed. Documents included scanned pages, rotated drawings, inconsistent layouts, and domain-specific terminology.
The pipeline combined machine learning with deterministic rules and validation logic. ML handled OCR, classification, entity recognition, and document understanding, while rules and scoring mechanisms helped constrain and verify the results.
In other words: OCR extracted the text; the processing system turned it into controlled engineering data.
Bring a representative sample of your documents, extraction requirements, and quality constraints. Azati can help determine whether standard OCR is sufficient or whether a custom document intelligence pipeline is justified.
Discuss your document-processing challengeAzati built a custom GPU OCR stack for high-resolution technical documents and P&IDs. Polygon-level detection helped handle text in orientations and layouts that conventional bounding-box approaches struggle with.
Document classification was implemented with a hierarchical DistilBERT-based model. Rather than treating every page as an isolated input, the model used the history and context of previous pages when classifying documents. This matters when a technical record consists of multiple pages whose meaning depends partly on where they appear in the document.
The pipeline also needed to extract technical entities rather than simply produce searchable text. Azati developed a custom DistilBERT-based architecture for NER and applied it to information relevant to the client's technical records workflows. The extraction layer was complemented by scoring and out-of-distribution detection using CatBoost.
Azati connected the individual AI components into a complete processing workflow:
PDF → image conversion → text detection → OCR → rotation correction → page classification → entity extraction → title extraction → QA/QC → reconciliation → structured reports.
The pipeline also produced searchable PDFs with embedded OCR text.
At approximately 5 million documents, the pipeline required distributed processing rather than isolated ML workloads. Azati used Ray, KubeRay, Kubernetes, and NVIDIA GPUs with asynchronous batching to coordinate large processing jobs.
The resulting system combined custom GPU OCR, hierarchical classification, domain-specific extraction, and operational controls into a repeatable distributed processing workflow.
| Delivery area | What Azati delivered | Business impact |
|---|---|---|
| Bulk document processing | Distributed Ray + Luigi pipeline across 15+ Ray jobs and 20+ processing steps | Created a repeatable processing workflow for a ~5M-document technical estate |
| Technical OCR | Custom GPU OCR with detection, recognition, rotation handling, tiled inference, and polygon-level processing | Made difficult engineering PDFs suitable for downstream classification and extraction |
| Document classification | Hierarchical BERT classification using page-history context | Added document/page structure to the processing workflow rather than treating pages as isolated images |
| Metadata extraction | Domain-specific NER and extraction for AFE, FTN, facility, line, country, and other technical information | Converted relevant technical-record content into structured data |
| AI quality controls | OOD detection, scoring, QA/QC generation, deterministic business rules, and reconciliation | Added control mechanisms around probabilistic extraction |
| PDF processing | Page/PDF de-rotation and OCR text embedding | Produced searchable versions of processed technical documents |
| Structured reporting | Four Excel report types covering indexes, document metadata, FTN metadata, and record metadata | Made extracted information available in structured operational formats |
| Data and task management | FoundationDB task tracking and ClickHouse reconciliation | Supported coordination and validation across the distributed processing workflow |
| Production infrastructure | Kubernetes, KubeRay, NVIDIA CUDA/GPU infrastructure, Docker, GitLab CI, DVC | Implemented infrastructure for repeatable distributed document processing |
Not sure whether your current OCR, document AI, or manual workflow can meet the required quality and throughput? Start with document types, extraction targets, validation rules, and a representative sample.
Plan a document AI assessmentThe architecture provided a repeatable workflow for OCR, classification, extraction, validation, and reporting.
The pipeline produced searchable PDFs and structured outputs containing document, metadata, FTN, and record information.
Confidence scoring, OOD detection, reconciliation, deterministic rules, and QA/QC provided mechanisms for identifying results requiring additional verification.
This type of engagement is best suited to large energy, Oil & Gas, utilities, engineering, infrastructure, and industrial organizations with substantial technical-record estates that need to become searchable, structured, and operationally usable.
Typical stakeholders include organizations responsible for engineering information management, asset data modernization, document digitization, or AI initiatives that depend on legacy engineering documentation.
It is particularly relevant when an organization:
A custom document intelligence pipeline becomes appropriate when standard OCR can recognize text but cannot reliably produce the structured, validated information the business needs.
Standard OCR may be sufficient when the requirement is searchable text from clean, consistent documents. Custom engineering becomes more relevant when the documents contain drawings, irregular layouts, domain-specific entities, multiple page types, or strict validation requirements.
For example, an organization may need to know not only what text appears in a document, but whether a page belongs to a particular record, which technical entity it references, what metadata should be associated with it, whether the extracted result is sufficiently reliable, and how that information should enter downstream workflows.
Oil & Gas companies can automate technical document processing by combining OCR, classification, metadata extraction, validation, and structured reporting in one workflow. The practical challenge is that technical records often require domain-specific rules and context rather than text recognition alone. In this engagement, Azati combined custom OCR, hierarchical document classification, domain-specific entity extraction, business rules, QA/QC, and distributed processing to process a technical document estate of approximately 5 million documents.
AI can extract metadata from technical documents when models are trained or configured for the document structures and entities relevant to the business. In practice, reliable metadata extraction usually requires a combination of machine learning, document context, deterministic rules, and validation rather than a generic language model alone. Azati implemented domain-specific extraction for information including AFE, FTN, facility, line, and country data, together with OOD detection, scoring, QA/QC, and reconciliation.
Technical document OCR has to handle factors such as drawings, rotated pages, complex layouts, technical terminology, and very large documents that can make conventional OCR insufficient. The practical implication is that OCR may need to be combined with layout-aware detection, rotation handling, tiled inference, and domain-specific downstream processing. Azati developed custom GPU OCR using text detection, recognition, rotation detection, polygon-level processing, and tiled inference for the client's technical records.
AI can classify document pages individually, but page-level classification can become more reliable when the model also considers document context. The practical implication is that sequential technical records may require page-history or surrounding-page information rather than independent classification of every page. Azati implemented hierarchical BERT-based page classification incorporating page-history context.
AI-extracted information can be validated by combining model confidence, out-of-distribution detection, deterministic business rules, reconciliation, and human or operational QA/QC where required. The practical implication is that probabilistic extraction should be surrounded by controls rather than treated as automatically correct. In this engagement, Azati combined CatBoost-based OOD detection and scoring with business rules, QA/QC generation, ClickHouse reconciliation, and structured reporting.
A custom pipeline becomes relevant when document volume, document complexity, domain-specific extraction requirements, or quality requirements exceed what a standard OCR product can reliably provide. The practical implication is that the decision should be based on the required business output, not simply on the number of documents. In this engagement, a previous approach had not achieved the required extraction quality, leading to a custom architecture combining GPU OCR, domain-specific ML, business rules, distributed processing, and QA/QC.
Document digitization generally focuses on converting physical or scanned information into a machine-readable representation, while document intelligence adds classification, extraction, contextual interpretation, validation, and structured outputs. The practical implication is that companies looking to use technical records operationally usually need more than searchable text. Azati's engagement went beyond OCR to classify pages, extract domain-specific entities and metadata, generate structured Excel reports, reconcile results, and produce searchable PDFs.
Cost depends on document volume, document types, scan quality, extraction requirements, validation thresholds, infrastructure, downstream integrations, and the degree of human review required. A proof of concept may use a limited document sample, while millions of technical records require distributed processing, model optimization, orchestration, storage, and operational controls.
Yes. A document intelligence pipeline can produce searchable PDFs, structured files, APIs, database records, or other outputs required by downstream engineering, asset, records-management, or reporting systems. Integration requirements should be defined alongside extraction requirements.
Explore the relevant projects and see how Azati delivers measurable results for the clients.
A major Middle Eastern oil and gas operator struggled with manually reviewing and reconciling massive volumes of engineering documents, leading to operational delays and errors. The client required a scalable, automated verification solution that integrated seamlessly with their existing engineering workflows.
Azati developed an AI-powered platform that automates the extraction and validation of technical documents and AutoCAD drawings, seamlessly integrating with the client's Knowledge Hub to streamline reviews without disrupting existing workflows.
An oil and gas operator struggled to utilize 35,000 static engineering drawings (PEFS) due to fragmented file formats (AutoCAD, PDF, TIFF), which hindered searchability, process tracing, and system integration. Manual reviews caused significant operational delays, creating a need for a scalable, automated solution to convert these documents into structured, machine-readable data while preserving engineering relationships and industry standards.
Azati’s AI-powered platform automates PEFS drawing processing, converts data to the DEXPI standard, and validates results for system integration. This transforms static archives into a searchable knowledge base that improves decision-making, streamlines workflows, and supports future digital twin initiatives.
A large oil and gas operator needed to extract and classify tag data from engineering drawings and identify equipment connections from wiring diagrams. Manual processing required substantial engineering effort, while tag meaning could vary by drawing type and connection data was encoded in drawing geometry rather than text.
Azati built an end-to-end computer vision pipeline that extracts text and symbols, classifies tags using visual context, reconstructs equipment connections from wiring diagrams, and generates structured AVEVA Engineering tables for engineer review. The production pipeline was delivered in three months and processes up to 300 documents per hour.
A large industrial operator needed to extract flange and engineering data from low-quality legacy P&IDs and piping isometrics to prepare Flange Master Registers (FMR). Degraded scans, inconsistent layouts, handwritten annotations, and ambiguous graphical elements made conventional OCR and fully autonomous AI unreliable.
Azati built a hybrid engineering document intelligence workflow combining computer vision, multimodal AI, deterministic validation, and human engineering QA. The solution identified flange joints, interpreted engineering annotations, structured the extracted data, and prepared validated FMR outputs while keeping engineers in control of ambiguous cases.
Engineering teams accumulated valuable knowledge across capital projects, but finding comparable historical observations required manual searches through project records, reviews, and documentation. The client needed faster knowledge retrieval without introducing a separate tool or disrupting established engineering workflows.
Azati embedded AI-powered semantic search directly into the client's existing capital project platform. The solution combines domain-restricted retrieval, LLM orchestration, contextual recommendations, and enterprise controls to help experts find relevant historical observations within their normal review workflow.
Last updated
