- Company
- dotData
- Industry
- Enterprise AI, AutoML platform
- Role
- Sr Product Designer
- Scope
- Research, prototyping, visualization, interaction design
- Team
- Product, Data Science
Making AutoML Self-Serve: From Fragmented Data Setup to Confident Model Runs
Summary
I led the end-to-end design of an AutoML setup experience that enabled Business Analysts to configure and run their first experiment with less reliance on Customer Support.
The work consolidated a fragmented, three-page workflow into a guided, single-surface flow that inferred data schemas, suggested table relationships, and made time-aware data configuration approachable through sensible defaults and just-in-time explanations. Automation stayed inspectable: users could validate, adjust, and rerun a configuration without starting over.
The context
Business Analysts were expected to use AutoML with limited machine-learning expertise, yet configuring a first model required navigating three disconnected workflows and often relying on Customer Support for guidance. The hardest parts were understanding how tables related to one another and defining time-aware data settings correctly.
To make AutoML truly self-serve, we needed to turn this expert-led setup process into a guided workflow that users could validate, adjust, and rerun with confidence.
Design challenge
Design a scalable AutoML configuration experience that enables non-technical Business Analysts to prepare data, validate table relationships, and configure time-aware prediction settings with clarity and control — turning an expert-led workflow into a confident path to a first model run.
The problem
Setting up a model was fragmented across three separate workflows: defining data sources and schemas, creating a use case, and configuring the model design task. Users had to manually connect tables, map fields, and configure time-based data logic — tasks that required machine-learning knowledge most Business Analysts did not have.
As a result, users often relied on Customer Support to configure or iterate on a model. Even when they could proceed independently, they lacked confidence that their table relationships and time settings were correct.
Before: importing data sources, selecting a target schema, and configuring data slots each lived in a separate place, so no screen showed how the setup fit together.
Key pain points
- Fragmented setup: critical configuration was spread across three pages, making the workflow difficult to follow.
- Hidden ML complexity: table relationships, data availability, and prediction timing were unfamiliar concepts for non-technical users.
- Limited confidence to iterate: users could run a model, but did not know what to adjust when they needed to rerun it.
Research and discovery
To understand where users got stuck, I interviewed Customer Support to learn about common customer use cases and recurring setup questions. I also partnered with the Head of Data Science to map the underlying ML workflow, and observed users configuring models to identify repetitive tasks, points of confusion, and moments that required expert intervention.
What I learned
Configuration knowledge lived with experts
Business Analysts could define a business goal, but often needed Customer Support or Data Science to translate it into valid data relationships and model settings.
Time-aware data setup lacked a clear mental model
Users did not know how prediction timing, relationship timing, and data availability affected a model, making them unsure whether their configuration was valid.
Automation needed to remain inspectable
Users wanted help connecting tables and applying defaults, but needed to validate and adjust recommendations rather than accept a black-box configuration.
Iteration should not require a restart
When a model needed to be rerun, users needed to adjust a specific setting and continue — not repeat the entire setup process.
Design strategy
I structured the redesigned experience around four principles: reduce unnecessary setup, preserve user control, explain complexity in context, and support iteration.
Four principles
Start with smart defaults
Infer data types and apply sensible defaults so users can begin configuring a model without needing to understand every technical parameter upfront.
Automate, then let users validate
Suggest table relationships automatically to reduce manual mapping, while allowing users to inspect, edit, and confirm every connection.
Explain complexity in context
Use inline explanations for advanced concepts — such as prediction timing, relationship timing, and data availability — only when users need to make a decision.
Design for iteration, not one-time setup
Keep configuration editable after a model run, so users can adjust inputs and rerun a model without rebuilding the workflow from scratch.
The solution: a guided AutoML workflow
The redesign consolidated three disconnected setup experiences into one guided workflow. Rather than asking users to understand the entire ML configuration model upfront, the interface moved from data preparation to model run through a sequence of focused decisions.
After: the same decisions on one surface — three separate setup workflows became a single flow users can move through, validate, and return to.
Upload and validate data
The system inferred data types on upload, allowing users to validate the schema instead of defining it manually. This reduced the initial setup burden while keeping users informed about how their data would be used.
Selecting already-uploaded tables or importing a new CSV. Each column’s type is inferred from the sample data and shown with its distribution, so users confirm the schema here — or open the table later to edit it.
Define the prediction goal
Users selected a target table, target column, and prediction time in one place. When table mapping was required, it appeared within the same step rather than as a separate workflow.
Target table, target column, prediction type, value mapping, entity ID, and prediction time in one panel — with “What is prediction time?” available inline, and the tables it refers to still visible on the canvas.
Connect tables with guided validation
An auto-connect action suggested table relationships based on the uploaded data. Users could inspect, edit, or remove each suggested connection, balancing automation with the control needed to trust the configuration.
Auto-connect proposes relationships from the uploaded data and reports what it added; hovering a connection shows the columns it joined on, with edit and remove beside them.
Validation names what is missing and offers the action that fixes it, instead of failing at run time.
Configure, run, and iterate
Time-aware settings appeared in context through defaults and inline explanations. After running a model, users could return to any configuration step, make a targeted adjustment, and rerun — without rebuilding the setup from scratch.
Time match, time range, and search range explained where the decision happens — with a default range inferred from the detected time column and an interactive timeline for manual setup.
The result: target, source tables, and their relationships stay visible while the model is configured, run, and adjusted.
Outcomes
What changed
- Reduced 20+ configuration actions to 5 guided steps.
- Supported faster iteration.
- Improved decision clarity.
My involvement
I led design across research, prototyping, visualization, and interaction design, partnering with product and data science teams to define a trustworthy, intuitive configuration feature.