PipelineGPT · Pipeline Integrity
Asking pipeline integrity data a question, and getting an answer you can trace
An AI interface that turns inspection reports, SCADA exports and PHMSA incident records into a queryable knowledge base, with a citation on every answer and an engineer in the loop.
Recognition
2026 Rosen Award Finalist
Type
Self-hosted AI platform
Languages
10
The answers exist, but they are buried in documents
Pipeline operators hold years of in-line inspection reports, SCADA exports, incident records and compliance schedules. Finding one fact across them can mean hours of reading, and it usually falls to the few people who know where to look.
General-purpose AI tools are a poor fit. In this field an unsourced or unreviewed recommendation is a safety risk, and operators are rightly cautious about sending integrity data to systems they do not control.
A natural language interface with safeguards built in
PipelineGPT lets operators and engineers ask questions about their own documents in plain language. The project is led at Origin by Collins Kubu.
Natural language queries
Ask a question with no SQL and no special syntax. Answers are retrieved from the operator's own uploaded reports and records.
Cited, traceable answers
Every answer cites the exact source passage: the document, the section and, for PDFs, the page.
Engineer review, always
High-risk recommendations are routed to a qualified engineer, who approves, edits or rejects them before the operator sees them.
From an uploaded document to a reviewed answer
01
Ingest
Operators upload PDFs, CSV files and PHMSA records. Documents are processed and indexed on the operator's own servers.
02
Retrieve
A question is matched against the indexed passages, and the most relevant ones are selected and reranked.
03
Answer
Claude writes an answer grounded in those passages, with a citation for each claim, in the language the question was asked in.
04
Review
Answers that contain an operational directive, such as a repair, shut-in, pressure reduction or evacuation, or that fall below a confidence threshold, are held for engineer sign-off.
Designed for operators who cannot take chances with their data
- Self-hosted: document processing and search stay on the operator's infrastructure
- Answers are written by Claude through the provider the operator chooses: Anthropic, AWS Bedrock or Google Vertex AI
- Every sign-in, question, upload and engineer decision is logged and exportable for audits
- Invite-only access with operator, engineer and admin roles
- Mandatory two-factor authentication for engineers and admins, and lockout after repeated failed sign-ins
- Open source under AGPL-3.0, so a security team can review every line that touches their data
A 2026 finalist for the Hermann Rosen Award for Pipeline Innovation
PipelineGPT was named one of three finalists for the 2026 Hermann Rosen Award for Pipeline Innovation, presented by the ASME Foundation. It was entered by Collins Kubu as "PipelineGPT: A Retrieval-Augmented Natural Language Interface for Pipeline Integrity Data".
Under the hood
See it for yourself
Working on something where the technology has to be dependable and the stakes are real? We would like to hear about it.