Bechtel’s project engineers had a backlog of roughly 130 AI use cases, but building any of them meant going through Databricks data scientists and engineers, a bottleneck that couldn’t scale. Bechtel wanted a governed way for its own engineers, without a data science background, to build AI workflows themselves.

Lovelytics designed and built BEC Nexus, a no-code, drag-and-drop workflow canvas on top of Databricks, so Bechtel’s engineers can assemble their own AI use cases, validated against an existing hand-built workflow during an eight-week engagement that followed an on-site workshop.

Bechtel builds some of the most complex projects in the world: data centers, mines, refineries. Its project managers and field engineers already relied on Databricks for their data and AI work, and they liked what the platform could do. What they didn’t have was a way for the people closest to the work to build their own AI use cases without waiting on specialized technical support.

Bechtel first asked whether existing Databricks agent tooling could solve this. It could, technically, but Bechtel wanted something more visual: a canvas where engineers could drag, drop, and wire together a workflow themselves, no code required. It was a hard enough problem that Lovelytics was initially hesitant to take it on, and even Databricks’ own product team weren’t sure it was possible.

Rather than commit to a long build with an uncertain outcome, both sides agreed to find out fast. Bechtel proposed a one-day, on-site workshop to see whether a working prototype was even possible; Lovelytics pushed to extend it to a day and a half.

In the two weeks before the workshop, the team researched front-end, back-end, and Databricks integration patterns that could support a true drag-and-drop experience, then spent the weekend before the workshop, including time on the flight out, building a working prototype: a functional canvas with an embedded AI assistant, ready before the team ever sat down at the table.

The prototype landed well enough that what was originally an eight-week discovery-only conversation turned into a real build. To keep the sequence straight: a weekend prototype, then the workshop, then an eight-week engagement in two phases, with one use case delivered at the end.

The first four weeks of that engagement went to the hard, unglamorous problem of making sure every action inside BEC Nexus ran under a user’s own Databricks permissions rather than a shared service account, so every workflow stayed auditable. 

The remaining weeks were spent building “Requirement AI,” the engagement’s one production use case, a use case that lets engineers parse project requirements out of PDFs and project plans automatically, validated head-to-head against a version Bechtel had already built directly on Databricks.

A second use case, a “work packaging” tool that combines 3D design data, material specs, and resource constraints to help plan how a project gets built, turned out to be complex enough that it made more sense as its own standalone application, still hosted on Databricks but outside the no-code canvas. A third proposed use case was set aside so the team could stay focused on the core platform within the eight weeks.

Bechtel valued a partner willing to say plainly when something was going to be difficult rather than overselling a safe answer, and was comfortable with Lovelytics using AI-assisted development to move quickly. That combination of candor and pace is what carried an uncertain discovery conversation into an extended engagement, with Bechtel adding more people and more time to keep building through the end of the year.

Not everything landed cleanly on the first try. A few weeks after real end users got their hands on BEC Nexus, their feedback came back as a long list of requested changes. Rather than treat that as a setback, the team used it as the basis for a full rebuild: refactoring the entire codebase overnight into modular pieces so four engineers could work in parallel without stepping on each other, tracked across roughly 90 work items and six parallel AI coding agents, with twice-daily syncs to resolve conflicts. 

One week went to building, the next to user acceptance testing, and the result was an application the team described as looking completely different, plus in-app onboarding, tutorials, and a use-case marketplace for end users.

Eight weeks after a workshop nobody was sure would produce anything usable (following a prototype built over the weekend before it), Bechtel had a governed, self-service AI platform its own engineers could build on, plus a validated playbook for rebuilding it fast when real user feedback demanded it.

• Working prototype built over the weekend before the workshop, ahead of the 8-week engagement
• 8-week engagement, starting after the workshop, took the platform from prototype to one validated, production-ready use case (Requirement AI)
• Full platform rebuild, driven by real user feedback, completed in 2 weeks using 6 parallel AI coding agents across roughly 90 tracked work items
• 130 AI use cases in Bechtel’s backlog now addressable through self-service, no-code builds
• In-app onboarding, tutorials, and a use-case marketplace now live for end users

Bechtel is now building an admin page to manage governance and user permissions across the platform, with 5 to 10 new use cases planned for the marketplace in the coming weeks. Lovelytics and Bechtel are also documenting the delivery methodology behind this engagement to apply on future projects.