AI Builder 0→1 · UX Research + Design Expert — San Francisco, CA
I turn user insight into shipped products.
Research-led. Design-driven. Deployment-ready.
9+ years bridging the gap between deep user understanding and products that actually ship — from ecosystem research at Google to enterprise CRM at ServiceNow to AI tools at Intuit to independent 0→1 builds using Claude, Cursor, and Vercel.
Hi, I'm Sonia. I turn deep user understanding into seamless product designs — that ship.
With 9+ years across Google, ServiceNow, and Intuit, I've studied how people interact, think, and the challenges they often encounter using digital products—spanning consumer, enterprise, employee tools, complex systems, and climate focused solutions.
I build end-to-end products from discovery to deployment.
Curious if we'd work well together? Send a message →↓
UX Research · Enterprise · AI
Role-Based AI Tools for CRM Adoption
Helping design AI-driven workflows that increased CRM adoption from 15% → 96% Avg. MoM (Month over Month).
Role
Lead UX Researcher
Scope
AI Outreach Assistant + Sales Intelligence Hub
Methods
Interviews, Survey, Co-design, ESAT
Company
ServiceNow
Context
The CRM platform had ~15% monthly engagement due to poor usability, fragmented workflows, and low information discoverability. Sales teams struggled to complete key tasks efficiently, leading to inconsistent data, low visibility for leadership, and missed revenue opportunities.
Core Insight
"Adoption fails when tools don't match real workflows. Sales users don't think in features — they think in workflows — and getting a job done."
CRM tools failed because they forced unnatural navigation instead of supporting real task flows. I approached this as a workflow and personalization problem, not a feature problem.
Approach
Method 1
Mapped workflows across 5 sales roles
Grounded the program in how each role actually works, not generic CRM tasks.
User Interviews · Focus Groups · Journey Mapping
Method 2
Identified role-specific data & task priorities
Pertaining relevance of data to surface the right information per role and reduce discoverability issues blocking progress.
Survey · Feature Stack Ranking
Method 3
Validated through co-design + post-launch research
Closed the loop with teams and users after shipped changes.
Co-design · FigJam · Information Widgets
Key Findings
After ~24 interviews across sales roles — Account Executives, SDRs, Sales Managers, RevOps, and Customer Success — a consistent pattern emerged: the CRM was failing not because people didn't want to use it, but because it didn't speak their language or serve their day.
Each role needed different data surfaced
An AE needs pipeline status and next-best-action. A Sales Manager needs team performance and forecast accuracy. The same CRM view served no one well.
Information was buried, not missing
The data existed in the CRM — but users couldn't find it without too many clicks. The AI layer's job was discoverability, not creation.
Trust in the tool was the real barrier
People defaulted to spreadsheets and Slack because the CRM felt unreliable. AI surfacing relevant, accurate, role-specific insights rebuilt that trust.
Solution
Helped design two AI-powered systems: an AI Outreach Assistant to streamline communication workflows, and a Sales Intelligence Hub to surface relevant data, alerts, and insights in real time. Both were role-personalized — adapting to how each sales function actually worked.
The use cases and interview feedback helped me work directly with engineers to design the AI models around real user needs — surfacing the relevant information people often needed but couldn't discover in a cluttered, one-size-fits-all CRM.
When workflows finally match how people actually work, the results speak for themselves.
The employee experience — finally a tool that works the way I do.
The leadership view — adoption numbers that moved the needle.
Reflection
Role-based personalization is critical in complex enterprise tools. AI becomes valuable when it surfaces the right information for the right person at the right moment — not when it adds more features.
Next Case Study
Health of Cross-Company User Goals
AI Research · Financial UX · Product Strategy
AI Avatars and Voice Assistance for the Right Task
Building a framework for when human-like AI or voice assistance adds value in high-stakes financial workflows.
Role
Lead UX Researcher
Team
AI Futures
Methods
Desk research, Usability & Evaluation of AI Avatar design vs Voice Assistant design
Company
Intuit
* Stock photo used for illustrative purposes. Not representative of actual study participants.
Context
Solopreneurs often manage finances alone and face moments of confusion, uncertainty, and emotional strain — especially when working through unfamiliar reports, fraud alerts, or high-consequence financial decisions. Leadership saw promise in AI avatars, but the core question wasn't whether avatars were possible — it was whether they would genuinely help users in the right moments.
Core Insight
"The real design question was not 'Should we use avatars?' but 'For which tasks do avatars or Voice assistance create meaningful value?'"
AI support is most effective when it aligns with the emotional and cognitive demands of the task. In moments of ambiguity, anxiety, or isolation, human-like guidance helps. In faster, lower-friction workflows, the same embodiment feels unnecessary or disruptive.
Approach
Method 1
Strategic desk research
Synthesized literature on conversational agents, anthropomorphism, and cognitive load to ground decisions in evidence, not novelty.
Method 2
Framework building
Created a Task-to-Avatar Value Framework to guide when avatar support appears and when other interaction modes are better suited.
Method 3
Prototype evaluation
Tested early avatar concepts to understand emotional response, trust, usability, and when embodied assistance felt helpful vs. intrusive.
Method 4
Cross-disciplinary synthesis
Drew on behavioral science (Nass & Reeves), HRI, cognitive load theory (Sweller), and multimedia learning (Mayer).
Solution
The outcome was not an avatar-first recommendation. It was a more disciplined product strategy: use avatars where users need reassurance, interpretation, or emotional support — and rely on chat or voice assistance when speed, flexibility, and control matter more. This shifted the direction toward a multimodal assistant model (option of Avatar + Voice assistance where either could be used) where assistance is invited, not imposed.
Impact
Direction Shifted
Product Strategy
Moved from avatar-first thinking to grounded multimodal strategy
Reusable Framework
Decision Tool
Evaluate avatar value by task type, emotion & cognitive demand
Trust-Centered
AI Design Principle
Control, timing, and modality matter as much as intelligence
The User at the Center
Solopreneurs managing finances alone — the exact user this research was designed to support. When AI assistance matches the emotional weight of the task, users feel understood, not overwhelmed.
The solopreneur user — navigating financial complexity alone. This research ensured AI support arrived at the right moment, in the right form.
Reflection
This work pushed the team beyond trend-driven thinking. Responsible AI design is not about making systems feel more human everywhere — it is about knowing when human-like support genuinely helps, and when it should step aside.
Next Case Study
Building My Own Portfolio
UX Research · Ecosystem · Product Strategy
Health of Cross-Company User Goals
Evaluating goal completion across the Google ecosystem — research that contributed to launching Google Family Link.
Google identified five high-priority ecosystem goals across Search, Chrome, Gmail, Google Account, and Assistant. The challenge: understand whether users could actually complete these goals across platforms — and where the ecosystem was failing them.
Core Insight
"Users don't think in products — they think in goals. The ecosystem requires people to know which product to use, creating a disconnect between user intent and product structure."
Approach
Method 1
Task success rate
Measured goal completion across all five journeys to identify where the ecosystem broke down.
Method 2
System Usability Scale (SUS)
Quantified perceived usability across products to compare journeys systematically.
Method 3
UX pitfall analysis
Catalogued friction points, confusion, and abandonment patterns per journey.
Key Finding
While most journeys showed partial success — users completing tasks with friction and confusion (yellow status) — parental supervision stood out as a critical failure. Parents were unable to discover or apply existing controls, making this the only journey that broke down entirely (red). This stark contrast highlighted a systemic gap and signaled an urgent need for a purpose-built solution.
This work touched something real — kids' safety. It reinforced that ecosystem design must start from user need, not product structure. The most impactful research isn't always the most complex — sometimes it's just being honest about where things break.
Next Case Study
AI Avatars and Voice Assistance for the Right Task
Product · 0→1 · AI-Assisted Build
Data Science App
Designing and shipping a data exploration tool for non-technical users — from research to deployed product.
Data tools are powerful but overwhelming — built for experts, not for people who simply want to explore and understand their data. Non-technical users struggle to upload data, generate insights, and interpret results. Most tools introduce complexity before value, causing users to abandon early.
Core Insight
"Users think in questions, not queries. People don't start with syntax — they start with intent. The system needs to meet them there."
Approach
01
Mental models vs. tool expectations
Mapped how people think about their data against what the product assumed they already knew.
02
Cognitive load reduction
Reduced friction across upload, exploration, and interpretation steps.
03
Progressive disclosure
Surfaced depth only when users were ready, instead of presenting upfront complexity.
04
AI-assisted build pipeline
Figma → Figma Make → Cursor → GitHub → Vercel. From insight to deployed product independently.
Solution
A guided exploration experience: Upload → Explore → Visualize → Interpret — no coding required. Clear drop zones, insight-focused UI prioritizing patterns and charts over raw metrics, structured information hierarchy that surfaces value before complexity.
Software Stack
✦Figma
✦Cursor
✦Claude Code
✦Railway
✦Docker
✦GitHub
✦Netlify
Designed layouts and interaction flows in Figma before touching code
Used Cursor + Claude Code to implement the full front-end independently
Containerized the backend with Docker and deployed via Railway
Shipped to production through GitHub → Netlify pipeline
Learned that prompt quality directly determines build quality
Gallery
Impact
End-to-End
Full Ownership
Research → design → code → deployment, independently
AI-assisted development is changing the role of designers. The ability to move from insight to product is becoming a core differentiator — and this project proved I can do it end-to-end.
Next Case Study
Role-Based AI Tools for CRM Adoption
Product Design · 0→1 · AI Matching
KidPool
Designing a hyperlocal AI-assisted carpool scheduler for busy parents — from concept to live product.
Role
Product Designer + Builder
Scope
0→1 Product, Workflow, IA, Interaction
Type
Product Design + Strategy
Status
Live ↗
Problem
Carpool coordination is still managed through reactive systems — group texts, WhatsApp threads, and spreadsheets. Overlapping schedules, constant changes, and trust requirements create high coordination friction and cognitive load for parents. There was no structured, trust-based way to express availability and get actionable matches without back-and-forth messaging overhead.
Approach
01
Defined the core coordination problem
Prioritized trust and clarity as the foundational design constraints — not just scheduling convenience.
02
Structured input over freeform text
Designed workflows that collect structured data to enable intelligent AI matching vs. reactive messaging.
03
Simulated scale with synthetic data
Stress-tested information structure and matching outputs before real users.
04
Built and shipped end-to-end
Took the product from concept → UX → implementation → deployment independently.
Product Screens
Activity Input Form — structured data collection for intelligent matching
Completed Activity Entry — clear visual feedback for parent inputs
Activity Confirmation Modal — immediate feedback with clear next steps
Intelligent Match Results — AI-powered recommendations with proximity data
Impact
End-to-End
Full Ownership
Concept → UX → implementation → deployment
Systems Thinking
Design Principle
Structured data enables matching — not just messaging
The most important design decision wasn't visual — it was insisting on structured input. That choice is what makes AI matching possible vs. just building another group chat.
Market Research · Growth Strategy · Personas
Expanding User Acquisition
Leveraging analytics and user research to refine personas and growth strategy — driving a 34% acquisition increase.
Role
UX Researcher
Scope
Personas, Analytics, Growth Strategy
Methods
30+ interviews, 200+ surveys, Usage analytics
Duration
~6 months
Problem
The team lacked clarity on who the right target users were and why potential users weren't adopting the product. Without grounded personas, acquisition messaging and onboarding were fragmented. Marketing needed data-backed personas and language for campaigns. Product needed insight into segment-specific goals and barriers to improve onboarding.
Approach
01
Quantitative usage analytics
Analyzed behavioral data to identify patterns in who was succeeding and where others dropped off.
02
30+ interviews with power users
Deep qualitative understanding of what drove adoption and value realization for high-engagement users.
03
200+ surveys with non-users & abandoners
Identified adoption barriers, mismatched expectations, and onboarding failure points at scale.
04
Synthesis into validated personas
Created distinct archetypes: analysts/power users, PMs/quick-insight seekers, marketing leaders/ROI-focused.
Solution
Translated validated personas into persona-specific messaging, onboarding improvements, and roadmap priorities. Gave marketing, product, and sales a single shared set of personas — eliminating the fragmentation that had caused misaligned campaigns and ineffective onboarding flows.
Impact
34%
User Acquisition
Increase over 6 months via persona-targeted messaging
52%
Onboarding Completion
Improvement via segment-specific flows and templates
3
Teams Aligned
Marketing, product & sales unified on a single set of personas
Reflection
The most powerful thing about this project wasn't the personas themselves — it was getting three teams to actually agree on who they were building for. That alignment unlocked the impact.
Design Engineering · AI Tools · Front-End
Building My Own Portfolio
A personal site built from scratch — where design thinking, prompt engineering, and code met for the first time.
Role
Designer & Developer
Scope
End-to-end
Tools
Figma, Cursor, Claude Code
Type
Personal Project
Context
Every designer has a portfolio. Not every designer builds one.
I started with Figma Make — designed the layouts, exported the HTML, brought it into Cursor. Version 1 went live. It worked. But something felt off. The fonts shifted. The animations felt heavy. Four case studies sat on the page, each structured differently, each telling a slightly different story about who I was. The inconsistency wasn't just visual. It was a thinking problem.
Core Insight
"Inconsistency across case studies wasn't a visual problem — it was a systems problem. Before fixing anything on the surface, I had to fix the structure underneath."
Approach
01
Audit before you build
Stopped adding. Looked at what was already there. Four case studies, four formats, no shared logic. Defined one repeatable template — problem, approach, tools, solution, reflection — and applied it across all four before touching a single line of visual code.
02
Prompt engineering as design tool
Moved from Cursor to Claude Code. Brought everything — every file, every section, every case study — into one folder. The ingredients were already there. What changed was the prompt. Bold. Editorial. Organized. Not cluttered. That single prompt, and the thinking behind it, unlocked the final direction.
03
Iterate with real users
Shared the site with people and watched how they read it. Not what they said — how their eyes moved. Adjusted hierarchy, spacing, and hover states based on where attention dropped. Every micro-interaction was a deliberate decision about mental models, not decoration.
Key Finding
Responsiveness exposed every weak decision hiding in the design. When fonts broke across breakpoints, it wasn't a code problem — it was a sign that the design system hadn't been thought through completely. Fixing responsiveness forced clarity. By the end, every element had a reason to be exactly where it was.
Tech Stack
✦Figma
✦Claude Code
✦GitHub
✦Netlify
✦HTML / CSS
✦21st.dev
Gallery
Drag the handle to compare before & after
‹›
Before
After
Impact
Live
Site Launched
Designed, built, and shipped without a team — start to finish.
3
AI Tools, 1 Workflow
Figma Make, Cursor, and Claude Code — each tool chosen deliberately, not by default.
4
Case Studies
Unified structure across all case studies, all devices, all form factors.
The site taught me something I hadn't expected. I already knew how to research. I already knew how to design. What I didn't know was how much I'd learn about both by having to build it myself. When you're the designer, the developer, and the user all at once — there's nowhere to hide.
