Studio
I rebuilt Walmart's agentic analytics platform in 4 months with 2 engineers, from a generator of fixed chart types into a workspace where one request returns a finished report.
- Role
- Product Manager III
- Dates
- March 2026 to present
- Team
- 2 engineers
- Users
- Every associate, from engineers to VPs
- Status
- In production, generally available
The problem
When I took over Studio, it was a proof of concept, not a product — something that worked in a demo but wasn't ready to hand to a business team, let alone get adoption. Underneath, it was guessing: which data source you meant, which model to use, based on whatever you happened to have access to. Guess wrong and you got the wrong table, ungoverned numbers, or a chart that didn't match the question.
Getting to that first guess took six setup steps and a separate login for every data source, and once you had something, there was no way to hand it to anyone else — sharing a dashboard meant copying the raw JSON behind it into a message.
What changed
| Before | After | |
|---|---|---|
| Getting started | Six setup steps before a first prompt | A workspace is ready on landing, so you type right away |
| Finding the data | Guessed from whatever you had access to | Routed to the authoritative source using catalog signals and session history |
| First output | Wrong tables, ungoverned numbers, mismatched charts | Shareable on the first pass, on-brand, with verified numbers |
| Editing | One edit usually changed every widget | Targeted edits, undo, redo, and one version history |
| Sharing | Copying the raw JSON into a message | A shareable link in three clicks, plus view-only chat |
Giving the agent something real to reason with
I didn't start by adding features — I pushed to delete the six steps between opening Studio and a first prompt, since they weren't protecting anything, just leftover scaffolding from an earlier version. Then I went after the guessing itself.
The fix wasn't a better guess. I originated an orchestration and memory model that reads catalog signals — what's certified, how often it's used, how people rate it, what access you'd need — against your own session history, so each question lands on the authoritative table. I shipped a lightweight version before the catalog filled in, because partial routing already beat guessing.
Guardrails that make output trustworthy
None of this matters if people can't trust what comes out the other end. I built a skill that validates a generated report against its real source of truth, and a skill that understands Walmart's retail calendar so a "this quarter" question doesn't quietly answer the wrong quarter.
The one I spent the most time on applies Walmart's design system to everything Studio generates. I tailored the system's primitive and semantic variables for data visualization, so every chart and layout takes its values from the system instead of the agent's guesses. I wrote it to make output accessible to everyone, not just on-brand. Then I A/B tested it, along with the validation skill, before either joined Studio's default agent.
Editing and sharing
In the version I inherited, editing one widget usually changed every other widget on the sheet, so a small fix could mean starting over. Telemetry showed regenerating a sheet cost about twice what an edit did. I restructured editing around targeted edits that change only what you point at, plus undo, redo, and a single version history. Each change now costs about 50% less.
Now a finished sheet reaches a shareable link in three clicks, with view-only chat, and one click publishes it as a governed marketplace source other teams can reuse. Smaller details mattered too: a sheet shows whether it's a draft or published with zero clicks, and every revision lives in one version list instead of a pile of separate copies. In launch week, 290+ people viewed shared reports.
What I didn't build, and where Studio lives
The single most requested feature, from analysts up through VPs, was real-time collaboration — and I didn't build it. I bet that a fast, trustworthy single-player experience mattered more than a slower multiplayer one, and held the people asking for it with publishing, version history, and a chat that already understands every number on a dashboard instead.
I set priorities from a telemetry dashboard I built inside Studio, tracking activation, retention, and token spend by surface, alongside user interviews, UX reviews, and competitive analysis against Power BI and Walmart's other internal tools.
The other big call was where Studio should live. I convinced leadership it needed to stand on its own as a product, rather than exist only inside Walmart's AI app platform, because that platform reached just one of Studio's user personas. Then I built that integration anyway, as an entry point to its 6K active builders. It drew 274 active users in its first week, 2.4× the prior month's total.
Shipping the front end, then teaching it
Getting Studio ready for every associate meant a lot of interface work, so I took over Studio's front-end repo and started shipping changes myself through Walmart's internal coding agent. I built the landing page that way, using Figma's MCP server so it came out compliant with Walmart's brand and design system.
Then I trained four PMs, all more senior than me, to clone the repo, branch, and ship front-end changes to their own products the same way.
Launch
Getting to production wasn't a feature-complete moment, it was a compliance one. Security certification had stalled for reasons that had nothing to do with the product being ready. I pushed it forward, made sure production infrastructure was built and waiting before the approval landed instead of after, and personally connected the people who needed to sign off on each piece.
Then I found the first users myself — cold-calling teams stuck building dashboards in other internal AI tools. Each became a repeat user who brought Studio to their own team, and an onboarding skill I wrote eased 10+ teams' move off those tools.
Looking back
Next time, the administrative steps — the security certification, the approvals — start on day one, not whenever the build has room for them; they gated the timeline more than any of the technical work did. The bigger lesson was about who Studio is for. Most tools like this only solve for the person building the dashboard. I built for the person just trying to read one, too — and that view-only mode ended up being one of the most-used parts of Studio, proof that the niche the existing tools left open mattered as much as the orchestration underneath it.
Studio is internal, so I don't show its screens here. I'm happy to walk through it live.