Overview
In this episode of The Dashboard Effect, Brick Thompson and Landon Oaks pull back the curtain on Celeste, an internal AI tool Blue Margin built to help data teams query, explore, and visualize client data using natural language. The conversation is candid about where the tool has succeeded, where adoption has fallen short of expectations, and what the team is learning in real time about what it actually takes to build an AI product that people use consistently.
It is a rare look at the messy middle of AI product development, told by the people doing the building. See how Blue Margin’s Managed Data Platform puts the infrastructure and AI tooling described in this episode to work for client organizations, turning the lessons learned building Celeste into practical capability for the businesses Blue Margin serves.
What This Episode Covers
What Celeste Does (0:31 – 1:23)
Celeste allows engineers to use natural language to generate SQL queries from metadata and semantic layers inside a data lakehouse. The design keeps sensitive client data out of external AI models while still giving teams a faster, more accessible way to extract and work with data. The security architecture is not an afterthought. It is central to how the tool was designed from the start.
A Fork in the Road (1:36 – 3:30)
The team is currently navigating a meaningful decision about who Celeste is really for. Engineers want workflow integration and efficient debugging capabilities. Business users want data visualization and immediate, interpretable insights. These are different products with different design requirements, and how Blue Margin resolves that tension will shape what Celeste becomes.
Adoption Lessons (2:20 – 2:50)
Early adoption among engineers was strong but tapered off as it became clear the tool did not always fit naturally into existing workflows. Brick and Landon are direct about what this revealed: building an AI product is an iterative process, and initial enthusiasm is not the same as sustained usefulness. The gap between the two is where the real product work happens.
What Comes Next (5:00 – 6:35)
The team is tracking developments across the AI landscape, including the release of GPT-5 and the potential of Azure AI Foundry, as inputs into where Celeste goes from here. The focus going forward is on agentic workflows and finding a more deliberate balance between the needs of technical and non-technical users.
Who It’s For
This episode is worth your time if you are a data engineer or BI professional curious about how AI tooling is being applied inside real data workflows, a product or technology leader building or evaluating internal AI tools and grappling with adoption challenges, a decision-maker assessing how natural language querying fits into your data strategy, or anyone interested in an honest account of what AI product development looks like beyond the launch announcement.
Why It’s Worth a Listen
Most conversations about AI tools focus on what they can do in ideal conditions. This one focuses on what happens after the initial excitement fades and the team has to figure out why adoption dropped off and what to do about it. That is a more useful conversation for anyone who has shipped an internal tool and watched usage plateau sooner than expected.
The fork in the road discussion is particularly worth attention. The tension between building for engineers versus building for business users is not unique to Celeste. It surfaces in nearly every internal data product, and the way a team resolves it determines whether the tool becomes infrastructure or shelfware. Hearing Brick and Landon think through it openly is instructive regardless of what you are building.
For organizations watching the AI landscape evolve and trying to make deliberate decisions about where to invest, this episode offers a grounded perspective from a team that is building in real time and learning as they go.
