Redesigning Seek AI
From convoluted SQL tool to conversational data assistant—in 4 months.
Impact at a Glance
• Acquired by IBM in 2025
• 4-month MVP timeline (design to deployment)
• Chat-first interface replaced complex SQL editor
• 4 user personas identified to expand the market from executives to entire data teams
• Verification system built to ensure data accuracy
Seek AI Version 1.0
The Challenge
Seek AI 1.0 had a problem: only SQL experts could use it.
The platform required users to write and edit SQL code to query their data. For business executives who understood data but weren't engineers, the interface was overwhelming. Users faced:
• Steep learning curve with unfamiliar features scattered across the interface
• No way to verify the accuracy of AI-generated SQL queries
• Complex onboarding to connect databases
• Difficult to identify which databases were connected or being queried
• No collaboration features for teams working together
The company needed to pivot from a SQL-focused tool to a natural language interface that anyone could use—and they needed it fast to pitch investors and potential clients like Snowflake.
Timeline: 4 months to design, build, and ship a complete redesign.
My Role
Product Designer & Manager (via AE.studio)
4 months | Led design, mentored junior designer
I joined to help the product team evaluate their design direction. Two weeks in, the Head of Product left. The CEO asked me to lead the entire redesign.
I worked in weekly sprints with a junior designer and a technical PM, conducting user interviews, developing the UX strategy, designing all screens, and managing handoffs to engineering—all while meeting aggressive investor deadlines.
redesigned home screen
The Insight That Changed Everything
Our users were already using ChatGPT.
In user interviews, everyone mentioned the same workflow: Ask ChatGPT to write SQL → Copy code → Paste into Seek → Hope it works.
They wanted Seek to feel like ChatGPT—a simple input bar where they could ask questions in plain English—no SQL required.
This insight shaped the entire redesign: Make Seek conversational, not transactional.
What I Built
1. Chat-First Interface
The Problem: Users had to navigate menus, understand SQL syntax, and manually structure queries.
What I Built:
• Conversational UI with input bar (similar to ChatGPT)
• Natural language queries → AI generates SQL → Returns results
• Eliminated the need to write or edit SQL code directly
• Made data accessible to non-technical users
Design Decision:
Users could still view/edit SQL if needed, but it was optional, not required. This preserved power-user functionality while making the tool approachable for everyone.
user personas
2. Four User Personas (Expanded Market)
The Problem: Seek 1.0 targeted SQL-savvy business executives. The natural language interface opened the door to a much broader audience.
What I Did: I conducted user interviews and analyzed V1 usage patterns to identify 4 distinct personas:
» The Data Engineer
Can write complex SQL, understands data structure, and answers technical questions.
Can write complex SQL, understands data structure, and answers technical questions.
» The Data-Literate Asker
Knows what data they want, can read some SQL, and formulates good questions.
Knows what data they want, can read some SQL, and formulates good questions.
» The SQL Writer
Reads and writes SQL, supports others in surfacing valuable data.
Reads and writes SQL, supports others in surfacing valuable data.
» The Simple Asker
Asks questions but doesn't know SQL. Needs guidance on how to frame queries.
Asks questions but doesn't know SQL. Needs guidance on how to frame queries.
Impact:
Expanding from 1 persona to 4 meant Seek could now serve entire data teams—not just executives—a key selling point for investors.
Expanding from 1 persona to 4 meant Seek could now serve entire data teams—not just executives—a key selling point for investors.
Seek workspace hierarchy
workspace management
3. Workspace Hierarchy & Permissions
The Problem: Multiple users across different organizations needed varying levels of access to proprietary data.
What I Built:
• Org Admin: Invite external users, manage database connections, promote other admins
• Workspace Admin: Invite org members, manage roles within workspaces
• Team Member: Edit/verify insights, update SQL
• Data Analyst: Chat with Seek AI, create workspaces, and add users to conversations
Reference Models:
I studied Slack, Microsoft Teams, and Google Workspace to understand how they structured permissions for shared proprietary information.
I studied Slack, Microsoft Teams, and Google Workspace to understand how they structured permissions for shared proprietary information.
Impact:
Users could now collaborate securely—sharing insights, verifying data accuracy, and working across teams without compromising data privacy.
insight page and verified/unverified screens
4. Verification System
The Problem: Users had no way to confirm if AI-generated answers were actually correct. Trust was low.
What I Built:
• Mark as Verified: Data analysts could confirm specific answers as accurate
• Verified badge: Highlighted trusted insights in chat
• Smart suggestions: Verified insights surfaced automatically for similar future queries
• Insights page: Centralized view of all verified/unverified answers with filters, search, and status tracking
Solved the biggest trust issue. Users could now confidently share data-backed answers with stakeholders, knowing data experts have reviewed them.
onboarding slides
5. Simplified Onboarding
The Problem: The complex setup process deterred new users.
What I Built:
• Onboarding slideshow introducing key features
• Option to connect real databases OR use a sample dataset for a test run
• Clear step-by-step flow for first-time users
Impact:
Reduced friction to get started. Users could explore Seek's capabilities immediately without committing to a full setup.
Defining the AI Agent's Personality
Since Seek AI wasn't just a tool—it was a conversational assistant—we needed to define how it should interact with users.
The AI Agent is:
• SQL Oracle → Writes simple to complex SQL code
• Data-Literate Disambiguator → Suggests better questions, helps explore databases
• SQL Copilot → Assists users in writing queries more efficiently
Personality Traits:
• Informative, professional, knowledgeable
• Confident in its abilities
• Patient and understanding (not all users know SQL)
• Friendly and approachable
I documented these characteristics for the engineering and LLM teams to guide agent development.
Interaction Design Details
Multi-User Chat
Challenge: When multiple users join a chat, who is Seek responding to?
Solution for MVP:
Users must @mention names to direct questions to them. @Seek triggers AI response; @user_name sends to human.
Users must @mention names to direct questions to them. @Seek triggers AI response; @user_name sends to human.
Future explorations:
• Pause Seek temporarily in group chats
• Side threads for specific insights
• Role-based visibility (e.g., @Seek responses only visible to data analysts)
Additional Features (Not Detailed Here)
• Data Connections: Connect data warehouses, upload CSVs
• SQL Editor & Schemas: Retained for power users
• Chat History & Activity: Track conversations and insight updates
• Data Tables & Charts: MVP included tables; charts deferred to V2
• Freemium vs. Paid: Defined feature tiers
• Live Human Support: Outsourced data analysts could join to assist
Seek 2.0 dark version
The Result
Shipped MVP in 4 months.
The redesigned Seek AI:
• Received positive investor feedback
• Helped secure partnerships (including Snowflake)
• Established the design direction still in use today
• Led to the IBM acquisition in 2025
