New: Chai, my AI assistant, can brief you on my work in 60 seconds.
Case study · Datagaps DataOps Suite · 2026
Designing a context-aware AI assistant that lives inside every page of the Datagaps DataOps Suite — answering product questions, writing validated SQL and starting real work, without pulling data teams out of their flow.
At a glance
The challenge
Power users manage hundreds of dataflows (one workspace holds 500 across 50 pages). Answers sat in docs, correct SQL depended on schema knowledge spread across people, and failures meant scanning run history page by page.
What I led
As Lead UX/UI & Product Designer for the DataOps Suite, I owned the assistant from strategy to the last component state.
The outcome
Methods
Pain-point mapping Jobs to starter prompts User groups Design principles Core flow mapping Information architecture State design Component kitFrom question to validated result in one panel.
01 Context
The DataOps Suite helps data teams test, validate and monitor data across sources, dataflows and pipelines. Power users manage hundreds of dataflows — the workspace below holds 500 across 50 pages — built from drag-and-drop components, SQL and rules. Getting work done meant juggling documentation, schema knowledge and run history in separate places.
Before — the Dataflows workspace: dense and capable, but silent when you get stuck.
02 The problem
How-to answers sit in docs and help articles. Users leave the canvas, search, and come back having lost their context.
A correct query depends on knowing the right datasource, schema, tables, keys and date columns — knowledge spread across people.
Understanding why dataflows failed means scanning run history across many pages of items, one by one.
Goals
How might we
…give data engineers and QA analysts an assistant that understands where they are, answers from trusted sources, writes SQL they can verify — and takes action, without leaving the screen?
03 My role
I led the assistant's experience from strategy to the last component state. I decided how it enters the product, how it talks, how it shows its work and how it fails gracefully, and designed it for the three groups who use the suite every day: data engineers, QA analysts and workspace admins.
I set the model: one entry point across every module, two specialised agents, and starter prompts mapped to the top jobs.
The core flow from question to validated result, clarifying choices, live reasoning, and side panel to full screen.
Built in the existing DataOps Suite language (Poppins and Inter, the Datagaps green accent).
Generating, thinking, not configured, history and delete. I turned them into a reusable kit so new agents can plug in without new UI.
04 Process
Map where data teams were losing time: docs outside the product, schema knowledge, failure tracing.
Three users, four design principles and one core flow from question to validated result.
Entry point, grounded answers, clarifying SQL, verifiable output, full screen, history, actions and empty states.
Twelve reusable components in the existing product language, ready for new agents.
05 Research
Looking at how power users worked across hundreds of dataflows, the same three jobs came up again and again. Each one became a starter prompt, so the assistant opens on the work people actually need to do.
| Job | What got in the way | Starter prompt |
|---|---|---|
| Learn how a feature works | How-to answers sit in docs and help articles, outside the canvas. | Summarize dataflow creation process |
| Write a correct query | The right datasource, schema, tables, keys and date columns are spread across people. | Generate SQL Query |
| Understand why runs failed | Run history is scanned across many pages of items, one by one. | Analyze recent dataflow failures |
06 Who it's for
Builds dataflows and pipelines
Needs correct SQL fast, and wants to see how it was derived before running it.
Writes rules, comparisons and validations
Needs quick, reliable answers about how product features work — in the moment.
Configures the suite for the team
Needs a clear path to connect an AI provider and govern how the assistant is used.
07 Design principles
A side panel opens over any page, so users never lose their place. It expands to full screen when a task needs room.
When a request is ambiguous — which datasource, which schema — the assistant asks with one-tap choices instead of guessing.
Live reasoning steps, detected entities, optimisation notes and a confidence score make every answer inspectable.
A Database Agent and a Document Agent scope each conversation, so answers come from the right source.
Core flow
Structure
Global AI is one entry point in the left navigation that opens on top of any module — Dataflows, Data Quality, Pipelines. Inside, an agent switcher in the composer scopes the conversation: the Database Agent for querying data, the Document Agent for files and documents. The thread lives in a side panel that can expand to full screen, with chat history alongside.
08 Key decisions
Getting work done meant juggling documentation, schema knowledge and run history in separate places.
One entry point in the left navigation that opens on top of any module, answers from trusted sources and starts real work.
The panel opens over the current page, so users never lose their place. Long answers and SQL reviews outgrow a 400px panel, so one click expands it to full screen — same thread, same state.
One tap to confirm a datasource costs far less than debugging a wrong query later.
Reasoning steps, confidence and a visible disclaimer did more to make answers usable than speed alone.
A Database Agent and a Document Agent scope each conversation, so answers come from the right source.
The assistant does not stop at text. It can build a Dataflow, narrating each step so users stay in control.
09 The solution
Global AI is one entry point in the left navigation that opens on top of any module — Dataflows, Data Quality, Pipelines. Here is how each part of the experience was designed.
The AI mark sits at the bottom of the global navigation and opens a side panel over the current page. A personal greeting and three starter prompts map directly to the top jobs we heard about.
Product questions return structured summaries drawn from official documentation and help articles — with clear Overview and Steps sections, embedded tutorial videos, and an offer to go deeper.
Instead of guessing, the Database Agent proposes the most-used datasource, asks how to handle multiple schemas, then asks what result the query should produce. While it works, each step is shown live.
The final query arrives in a syntax-highlighted, copyable code block — followed by the reasoning a reviewer would ask for, and a clear prompt for what to do next.
Long answers and SQL reviews outgrow a 400px panel. One click expands the conversation to full screen — same thread, same state — and collapses back to the side bar when the user returns to building.
Useful threads become reference material, so history is searchable and manageable in place — with a guarded, explicit confirmation before anything is deleted.
The assistant does not stop at text. Asked to “Create a Dataflow”, it analyses recent workspace activity, detects the preferred environment and engine, then builds the structure and source components — narrating each step so users stay in control.
The unglamorous states make a feature feel finished. Before an AI provider is connected, the panel explains what is missing and gives admins a direct path to configure it.
Panel moments
Building blocks
Each piece was built as a reusable component in the existing DataOps Suite language (Poppins and Inter, the Datagaps green accent) so new agents can plug in without new UI.
10 Validation
An AI answer is only useful if people can tell whether to trust it. Checking is designed into the flow itself, and every response carries a way to say whether it helped.
Live steps, detected entities and join keys show how a query was derived before anyone runs it.
Optimisation notes, execution complexity and a confidence chip sit right under the code block.
Response actions on every answer, with “Datagaps AI can make mistakes” visible beneath every input.
“Your query is ready and validated. What would you like to do next?”
11 What we shipped
entry point across every module
specialised agents — Database & Document
starter jobs mapped to real pain points
screens and states designed
12 Reflection
Reasoning steps, confidence and a visible disclaimer did more to make answers usable than speed alone.
One tap to confirm a datasource costs far less than debugging a wrong query later.
Not-configured, generating, history and delete states are what make an AI feature feel like part of the product.
