May to October 2025
From templates to AI agents
I turned a template-gallery assignment into an AI agent that builds machine learning pipelines from plain English and still gives experts a path back to code.
- Role
- Senior Product Designer
- Team
- 12 developers and 1 product manager
- Duration
- 5 months
- Scope
- Strategy, working prototype, and production design
- 17Pipeline step types
- Each required detailed configuration and connections
- 4Agent tools prototyped
- Generation, editing, configuration, and validation
- 3 moPrototype to production design
- Including user and leadership validation
- 50%Expected faster creation
- Post-launch estimate, not a shipped result
Replace a library with a system
- Problem
- Leadership asked for a gallery of static pipeline templates, but customers needed unique combinations and every template would create an ongoing maintenance burden for AWS.
- My decision
- I made the operational case for an AI agent that composes SageMaker's 17 step types from plain language, then proved the direction with a working prototype and visible pipeline canvas.
- Tradeoff
- A high-level chat experience made pipeline creation faster but hid details that advanced users still needed. I paired it with notebook export so people could move from intent to full control.
- Project outcome
- Leadership approved the AI-first direction, the prototype validated technical feasibility, and production designs were completed for engineering implementation.
Start with the direction that did not scale
The initial gallery made common pipelines approachable, but research exposed the ceiling quickly. A small change created a new template, a new maintenance obligation, or a fork customers had to own.
Let intent shape the pipeline
Ask before acting
The agent clarified data, scale, and performance needs before choosing steps. That made its reasoning inspectable instead of jumping straight to an opaque answer.
Build the workflow in public
The canvas updated as the agent added steps, connections, and configuration. People could see what changed and inspect the pipeline instead of trusting a hidden generation step.
Keep the path back to code
Chat covered the fast starting path. Notebook export preserved every parameter for experts who needed to refine algorithms, infrastructure, or production logic directly.
