Role Lead Product Designer
Year 2025
Team Founder & CEO, 1 Eng Lead, 2 Backend Engineers

From Data Discovery to AI-assisted Action

Designed Select Star's first AI agent — an auto-tagging system that moved users from finding data to acting on it. The agent processed 22K+ queries and governed 12K+ assets, delivering 6× engagement over legacy catalog workflows.

22K+ Ask AI queries
processed
12K+ Assets governed
through auto-tagging
Engagement vs.
legacy catalog workflows

The shift: from understanding data to acting on it

Select Star had built a strong foundation in data discovery — search, catalog, and lineage helped teams understand their data. But understanding wasn't enough. Customers like Amplitude were asking for tagging to "just work" automatically. The real gap was the last mile: moving from insight to action.

With no dedicated PM and a team already stretched across the context layer, AI settings, and LLM search, I had to drive alignment quickly. I wrote a lightweight PRD and mockups to get the CEO and engineering aligned on v0 and v1 before a line of code was written.

Find, understand Act

From helping users understand data context — to helping them act on it through AI

Before
Search · catalog · lineage
Now
An AI agent that does the work
Why auto-tagging first
01
Clearest customer pull
Customers expected it to "just work." The clearest unmet ask.
02
Built on our moat
Reuses lineage, metadata & Ask AI — least new build, most leverage.
03
Foundation for all agents
The rule engine becomes the base for every agent that follows.

Three directions explored — one chosen for the right reason

The core design challenge was trust: AI would be taking bulk action on enterprise data assets. Users needed to feel in control, not anxious. I explored three directions before landing on the final approach.

Option 01
Structured Wizard

Predictable and validated at each step, but too rigid for messy real-world instructions like "tag if PII and popularity > 80."

Rejected — too rigid
Option 02
Pure Chat

Maximum flexibility and fast to ship, but no way to preview what AI would do before touching data — too opaque for governance.

Rejected — too opaque
Option 03
Hybrid Chat + Plan Preview

Natural language input + structured preview before execution. Users inspect AI interpretation, see affected assets, and confirm before anything is applied.

Chosen — flexible input, structured verification
The design principle
The agent does the work. The human makes the call.

Designing guardrails for AI autonomy

The biggest design challenge wasn't the interface — it was building appropriate trust. Users needed intervention points before, during, and after AI workflow execution.

01 Plan & Confirm — turn natural language into explicit scope

Users type intent in plain language. The agent surfaces its interpretation + a preview of matched assets before anything is applied — AI logic inspectable, not a black box.

02 Streaming feedback — show users what the AI is doing in real time

The agent shows its reasoning live — interpreting → scanning → previewing — so users understand what's happening and why assets are included.

03 Turning AI failure into a recoverable workflow

Low ratings (<3 stars) trigger an admin review loop — AI diagnoses gaps, surfaces fixes for approval. Failure becomes a feedback signal, not a dead end.

From AI answers to governed workflows

Product Impact
22K+ Ask AI queries processed
12K+ assets governed through auto-tagging
6× engagement vs. legacy catalog workflows
Design Impact
Made AI logic inspectable and editable — not a black box
Created a reusable Plan & Confirm pattern adopted across future agents
Turned AI failure into a structured review workflow

The agent platform became one of Select Star's key differentiators in late-stage sales conversations — and a contributing factor to the company's acquisition by Snowflake in December 2025.