Phase 1
Discovery
Three weeks shadowing reviewers and cataloguing the top failure modes. We picked the eight clause categories responsible for most of the rework and scoped the pilot around them.
A top-five upstream operator wanted fewer surprises at FEED handover. We built an AI-assisted review layer that reads drawings, checks them against the standards, and tells reviewers exactly where to look.
A top-five upstream operator runs multiple capital projects in parallel, each producing thousands of P&IDs, isometrics, datasheets, and equipment specs that must be cross-checked against API, ASME, and a thick stack of internal engineering practices. The work was being done by senior engineers who were retiring faster than they could be replaced, and every project was sliding right by two to four weeks in the review gate.
The downstream cost was worse than the calendar slip. Issues caught after IFC — a missing relief path here, a wrong material class there — were turning into change orders during fabrication, with the usual multiplier on cost and schedule. Internal audits showed the same handful of clause categories showing up over and over.
The team had tried a rules engine. It found the trivial things and missed everything that required judgement. They wanted something that read like a junior reviewer who had already memorized every standard on the shelf.
Every finding cites the clause, the drawing region, and the prior project where it was last seen. Reviewers stay in charge; the system handles the page-flipping.
Vision pipeline that reads P&IDs and isometrics into a structured graph — equipment, lines, instruments, and tie-ins, with the title-block metadata attached.
Retrieval index over API, ASME, NACE, and the client's internal engineering practices. Every clause is chunked at section level and version-tracked.
Cross-checks line specs against datasheets, valve schedules against P&IDs, and equipment tags across discipline packages. Flags mismatches with a confidence score.
Side-by-side drawing viewer with a findings panel. One click jumps to the cited clause; another marks the finding accepted, deferred, or false positive.
Accepted findings become tickets in the engineering change workflow, pre-filled with the discipline, the drawing, and the clause text.
Every reviewer disposition feeds the next model release. The system gets quieter and more precise project over project.
Three weeks shadowing reviewers and cataloguing the top failure modes. We picked the eight clause categories responsible for most of the rework and scoped the pilot around them.
One brownfield package, one greenfield package. We ran the assistant in shadow mode for six weeks before any reviewer was asked to use it — so the false-positive rate was already tuned when it went live.
SSO, audit logging, role-based access, and the JIRA bridge. Performance work on the drawing parser so a 200-page package returns findings inside an hour.
Two additional project teams, a quarterly model release cadence, and an evaluation harness the client's own engineering team now runs without us.
Six months after go-live, the operator's program management office reported that engineering review had moved from the top schedule risk on every project tracker to the middle of the pack. Reviewers stopped opening PDFs of standards on a second monitor; the assistant surfaced the relevant clause text inline, with the right edition cited.
Just as important: the kinds of findings that used to surface during fabrication walkdowns — wrong material class on a relief tie-in, missing inspection requirement, an isometric that disagreed with its line list — were being caught in IFR. The change-order log on the first program after go-live was the cleanest the client had recorded in five years.
After the first program landed, the client expanded the contract to cover two more capital projects and a midstream pipeline integrity workstream. The reference is available; we just need a call first.
Talk to a reference →"We didn't want a tool that replaced our reviewers. We wanted a tool that gave them their afternoon back. East Reach built exactly that — and the auditors actually like it, because every flag points to a clause."
Reference architecture, evaluation methodology, and a redacted sample of the findings dashboard — sent over after a short call.