The challenge
A major energy operator faced thousands of operational and compliance documents daily across exploration, production, health and safety, and asset management, including drilling reports, well integrity assessments, environmental compliance records, safety reports and regulatory filings. Manual review created chronic backlogs, pushed turnaround on critical safety reports beyond regulatory SLAs, and produced inconsistent review depending on which analyst was assigned. The operator needed to automate the routine majority of document interpretation, roughly 80 percent, while keeping human oversight on the highest-risk 20 percent, with full auditability and reproducibility for regulators.
What we did
Designed a Databricks-based, LangGraph-orchestrated multi-agent workflow: specialised parsing agents for each document type (spreadsheets, PDFs, reports, email attachments), each tuned to its expected schema and anomaly patterns; a Delta Lake ingestion layer standardising extracted records into a governed canonical format; an agent evaluation framework scoring every output against accuracy and compliance benchmarks, with DeepEval and MLflow Evaluate providing continuous reliability tracking; and Unity Catalog lineage tracking the full audit trail from source document through model version to evaluation score. A modular orchestration pattern let the same architecture scale across use cases rather than being rebuilt per document type.
The outcome
Reduced manual data-validation time across daily and weekly compliance workflows, improved report accuracy with complete audit trails, and enabled near real-time operational insight during shift handovers and daily reviews. Established a reusable evaluation framework since applied to all of the operator's subsequent generative AI initiatives.
Automated the routine 80 percent of document interpretation, kept humans in control of the highest-risk 20 percent, and made every output traceable from source document to evaluation score.
Specialised agents parsed each document type into a governed canonical schema on Delta Lake, with an agent evaluation framework and DeepEval/MLflow Evaluate scoring every output against accuracy and compliance benchmarks. Unity Catalog lineage connected each output back through model version to source document, and the same orchestration pattern was designed to extend to new document types without a rebuild.