A 0→1 enterprise AI & data platform

The problem

Enterprise data teams run on a patchwork of partial tools. Storage, integration, governance, cataloging, and AI come from separate products that rarely talk to each other, so the work fragments, technical debt mounts, and no one gets a trustworthy view of how data flows or who can touch it. Engineers fight friction. Stewards fight blind spots. Executives can't get a straight answer about lineage.

The opportunity was a single platform where any role can work across the whole data journey on one auditable view of how things connect: engineer, analyst, steward, or executive.

How I led it

This was a five-year 0→1 initiative, and I co-led it with Product Management from concept to internal preview. I championed the research that justified the bet, set the design direction once the findings were in, and held both together as the team and the scope changed around them.

The group was cross-disciplinary: five researchers and consultants on the research program, two product managers, two engineers, and a design intern I mentored. Most did not report to me. Leading by influence, I set the principles we designed against, ran the brainstorms and reviews that shaped the project's direction, and kept overlapping user stories pointed at one coherent product.

At the beginning of the project, modern LLMs and coding harnesses weren't available yet, so I relied on a lean UX approach to prototyping in Figma. Midway through the project, however, that changed. In the AI-native era building got cheap enough that I did not have to ration it, so I leaned heavily into prototyping in code. We built an MVP to prove the concept worked as a whole, then kept prototyping in code well past it, using fast, disposable builds to shape the North Star vision and put those ideas in front of real customers and users. The MVP showed the whole could hold together. The prototyping is how we found what the North Star should point at, and built confidence in it before anyone committed.

The research I directed

I championed and led the research program that grounded the product, directing a five-person research and consulting team through two studies and 76 in-depth interviews. We talked with data professionals from hands-on engineers up to the C-suite, across healthcare, health insurance, financial services, and systems integration at large enterprises.

The research led to two major insights. The first: people shift between five mindsets as the work changes (building, transforming, analyzing, maintaining, orchestrating), regardless of their job title. An engineer titled “data engineer” might spend an afternoon analyzing and a morning orchestrating, and the title is a poor map of the actual work. We built a jobs-to-be-done model around the mindsets and a set of UX principles to design for the task, not the title.

The second: two pains were near-universal across every role and every industry we studied. Governance (“who can touch what, and how do we prove it?”) and data quality (“is this number right, and how would I know?”) came up in nearly every conversation. The clearest signal of all was that a visual, pipeline-centered view belonged at the heart of the product.

From research to design

I translated the research into UX principles, an end-to-end MVP (the minimum end-to-end workflow we would design against), and a North Star vision for where the product could go over time. The MVP proved the concept end to end. The North Star kept ambition.

Pipeline as the navigational hub

The central design decision was to make the pipeline itself the navigational hub of the product. Most data tools organize by feature: here are the connectors, here are the dashboards, here is the catalog. That structure forces users to context-switch every time the work moves between mindsets. We anchored the experience to the pipeline instead. Every resource (datasets, models, dashboards, AI/ML, code, tasks) is a composable node on the canvas, and lineage flows through it unbroken. Roles work on the same canvas with different mindsets. The canvas adapts. The data stays put.

No-code, low-code, pro-code on the same data

On the same pipeline, an analyst can work no-code (pick a transform, configure it through chips), an engineer can drop into low-code (edit the SQL the chips compile to), and a data scientist can go pro-code (a Python step on the same node). The handoff between roles is the same canvas, same data, same lineage. No export, no copy, nothing to re-run on the other side.

AI threaded through, not bolted on

A persistent AI assistant builds pipelines from natural language, but AI can also be added to the pipelines themselves as a step like any other element. This gives it a full lineage from prompt to execution and human-in-the-loop decisioning at each step. The design keeps AI inside the pipeline metaphor, so AI workflows are inspectable and auditable rather than a black box. In regulated environments like finance and health, that helps de-mystify workflows and keep decision making accountable.

A coded prototype to pressure-test the interactions

I coded a front-end prototype to pressure-test the canvas interactions and clarify the look and feel. The Figma files served as a stable reference to visual designs, but the prototype let me work out motion and rhythm beyond what static frames could communicate.

What testing told us

Demand was strong and consistent across every role and industry we tested. Participants repeatedly described the concept as long-awaited, and the pipeline-as-hub pattern tested best. The doubts were just as valuable. A recurring “if it works” skepticism about readiness, interoperability, and security directly shaped the roadmap and the design's emphasis on trust, traceability, and progressive disclosure of complexity.

By the numbers

5 yrs

Led design & research, 2021–2025

76

In-depth interviews led

10

Cross-disciplinary team led

Status

The platform is now being tested internally and shared with select advance-preview users. It grew out of the earlier SQL engineer experience and data pipelines work, and together the three trace an arc from feature to platform.