Overview

Hiring great talent is hard — especially for managers without a recruiting background. At LinkedIn, we set out to reimagine the hiring experience by designing an AI-powered assistant that helps managers write better job descriptions, source stronger candidates, and streamline the hiring process — all within LinkedIn Recruiter.

I helped lead product design from initial vision through launch, working across design strategy, prototyping, research, and delivery — as well as presenting to executives and completing final QA. The project earned the LinkedIn Company Innovation Award.

The problem

LinkedIn Recruiter was experiencing inconsistent and inefficient hiring workflows, which impacted non-recruiter hiring managers by increasing time-to-post, lowering confidence in candidate quality, and reducing engagement with advanced tools.

This mattered because hiring friction directly influences job posting revenue, customer retention, and perceived platform value.

Our goal was to design an AI-powered assistant that reduced cognitive load, accelerated decision-making, and increased hiring confidence — while preserving human control and aligning with LinkedIn's broader AI strategy.

The solution

We designed "Hiring Assistant," a smart layer inside LinkedIn Recruiter that helps managers post jobs, discover talent, and evaluate candidates — all with help from AI. The experience was rolled out in two phases: first, a lightweight posting and sourcing flow; then a more refined, customizable workflow based on beta feedback.

Vision

In early workshops, our team generated dozens of ideas to reimagine the hiring journey with AI. We aligned around a vision that focused on simplicity, clarity, and confidence — empowering managers, not replacing them. To gain stakeholder buy-in, we created high-level concept prototypes showing the end-to-end assistant experience.

Phase 1

To validate value quickly, I advocated for a constrained MVP focused on three key capabilities:

• AI-generated job drafts — AI generates a draft job post based on simple prompts. Managers can edit and adjust before publishing
• Automated candidate sourcing — the assistant finds qualified candidates to apply and pre-sorts applicants by the criteria you set
• Ranked applicant screening — as applications come in, Hiring Assistant evaluates and sorts top matches. Managers can save or reject in one click

This allowed us to test behavioral shifts without over-investing in unproven workflows.

Research & findings

During our "Friends & Family" closed beta, we gathered qualitative and quantitative feedback from high-value customers. These findings directly informed our Phase 2 design updates:

• Users wanted more manual control — users still weren't sure about the accuracy of AI performing tasks for them, so they preferred manual control or at least easier ways to review
• Job reposting needed to be easier — we needed to provide pre-written job descriptions or access to previously posted jobs to duplicate content
• Ranking credibility was a critical gap — nearly 80% of applicants were labeled "Top Match," diluting signal strength and reducing trust. This was a product-design-data systems issue, not just a UI problem

Phase 2

User feedback revealed a strong desire for more control and transparency. We pivoted from automation-first to collaboration-first:

• Guided review flow — we replaced 1-click publishing with a guided review process, increasing perceived agency and reducing anxiety around AI outputs
• "Why this candidate?" reasoning — we introduced reasoning modules to surface alignment between qualifications and role criteria
• Refined candidate matching — I partnered with data science to refine weighting logic, prioritizing quality over volume and reducing false positives in the "Top Match" category

This moved the experience from black-box automation to visible collaboration.

Design system contribution

As we were developing this product, other agentic products were being built across the company. We were all working in tandem and eventually consolidated to work together in a new "agentic" design system. I helped develop this new design system to work across the company — defining reusable AI behavior patterns, establishing emerging agentic standards, and aligning components with LinkedIn's design system to reduce downstream fragmentation.

Project outcomes

• 57% adoption rate among target users
• 6% lift in job posting revenue, quarter over quarter
• $230K in new revenue from the Promoted Plus tier
• Recipient of the LinkedIn Company Innovation Award

Landing page — Phase 1 vs Phase 2:

Candidate experience — Phase 1 vs Phase 2:

Hiring plan — Phase 1 vs Phase 2:

Learnings & challenges

Trust is still an issue with AI — transparency features like "Why this candidate?" increased trust. Time-saving features were a clear value driver, with managers appreciating better candidate matches.

Cross-platform fragmentation — eng and product teams didn't always work cross-platform, and most features were developed at different speeds, fidelities, and had different launch schedules. It was challenging to corral these as a platform-agnostic designer.

Innovation vs. shipping — in order to launch on time and find a happy medium with other parts of the product, we were limited in what type of innovative ideas and experiences we were able to ship.

Reflection

Hiring Assistant marked a strategic inflection point: defining how AI could responsibly augment professional workflows at scale. The future of AI in hiring isn't replacement — it's structured augmentation. This project operationalized that principle across product, platform, and business outcomes.

Overview

Tax experts spend most of their day supporting customers who often have complicated questions and emotional financial concerns. We saw an opportunity to ease that cognitive load by integrating Intuit Assist directly into expert workflows. The goal was to help tax experts work more efficiently while building stronger relationships with customers.

I helped lead product design from initial vision through launch, working across design strategy, prototyping, research, and delivery — as well as presenting to executives and completing final QA.

Defining the problem

Research with experts surfaced three core challenges:

• Time consumption — customer conversations were taking longer as issues became more complex
• Routine work slowed momentum — note-taking, summaries, and follow-up emails consumed a large portion of each shift
• Gaps in communication skills — experts often mastered tax language, but many wanted support writing with clarity, empathy, and a more personal tone

Project goals

Our objective was to automate repetitive tasks, improve access to contextual information, and support experts in providing thoughtful communication. We framed our design direction around a set of guiding questions:

• How can AI enhance expert workflows without disrupting control?
• How do we create more personalized experiences for customers?
• What information can we surface at the right moment to reduce cognitive load?
• How do we help experts communicate with confidence and clarity?

Research & discovery

We began with moderated studies focused on watching customer and expert interactions in real time. We paired that work with interviews that captured daily frustrations and opportunities. Privacy requirements were strict, so we used anonymized journeys and aggregated insights — preserving user trust while giving us enough signal to move with confidence.

This work required strong alignment across Intuit Assist and the broader AI organization. We partnered closely with engineering and product leadership to ensure our solutions could scale across products and future capabilities.

Mapping the expert journey

Experts interact with customers across text, email, video, and calls — each channel introducing a different workflow. For the first release, we focused on customer support calls and divided the experience into three phases: before the call, during the call, and after the call. Each phase contains tasks that require time and attention. By introducing AI support into these moments, we aimed to reduce effort and keep experts focused on the conversation.

Design solutions

1. A consistent entry point

Experts needed access to Intuit Assist without sacrificing customer data on screen. After iterative testing, we placed a small entry button in the global header — preserving space for high-value information while keeping AI support available at any moment.

2. Pre-call summaries

Intuit Assist automatically compiles essential customer details before an expert joins a call. This reduces time spent searching through records and helps experts prepare mentally for the interaction.

3. Natural language questions

Experts can ask questions in plain language and receive instant answers. Early testing showed an average time savings of twelve minutes per call because experts no longer switched tabs or conducted manual searches.

4. In-call summaries

Experts no longer need to take extensive notes. Intuit Assist monitors the conversation, captures key information, and creates structured documentation — leading to better continuity across future interactions.

5. Recommended responses

Intuit Assist suggests polished reply options that experts can edit to match their tone and personality. Recommendations are always optional — important for maintaining trust and autonomy.

6. Feedback inside the product

Experts can react to AI output with thumbs up or thumbs down controls. This influences future improvements and helps refine the model over time.

7. AI-drafted follow-up emails

Intuit Assist drafts a structured follow-up email based on the call. Experts can adjust tone and length before sending — removing the effort of recalling key points and rewriting the same information.

8. Onboarding & first-time use

To reduce friction, we introduced an animated onboarding moment that highlights the most helpful capabilities. We continued to experiment with guided tours and in-context prompts to improve feature discovery.

User testing & results

• Less context switching — experts saved measurable time by avoiding tab changes
• Clearer communication — generated summaries helped experts produce stronger written responses
• Control and trust — experts wanted to review and edit content before sending, confirming the importance of customization
• Better visibility after calls — experts asked for clearer indicators when summaries were ready

What we learned

Customization improves adoption. Early adopters were enthusiastic about efficiency gains, and editing and approval controls increased trust in generated output. Based on follow-up testing, we added more editing controls for generated content and introduced quick-select options for training and faster adjustments.

Project outcomes

• 12-point increase in NPS
• 15-minute reduction in average help times
• 88% expert engagement rate with Intuit Assist
• 28% reduction in silence or hold time during calls

Reflection

Integrating Intuit Assist into expert workflows improved efficiency, clarity, and customer connection. By focusing on real problems, validating assumptions through research, and iterating with intention, we created an experience that helps experts do their best work while supporting meaningful customer relationships. The key lesson: design the AI as an assistant that amplifies human judgment, not one that replaces it.

Overview

Facebook offers several profile types that allow people and organizations to represent themselves — personal profiles, additional profiles, professional profiles, and community profiles. However, users often struggled to understand the differences between these entities, especially between Personal Profiles and Professional Profiles (formerly Pages).

We also needed to meet legal requirements for displaying when a profile represents a business entity. I helped lead product design from initial vision through launch, working across design strategy, prototyping, research, and delivery.

Background

Our privacy and integrity partners asked us to improve how we surface two key aspects of an entity:

• Users should understand what type of profile they are viewing because permissions and capabilities differ
• Users should have high-level visibility into important integrity information about that entity

These requirements applied to all profile types and needed to remain discoverable without creating clutter.

Problems

User pain points:

• Difficulty distinguishing profile types
• Profile details were cluttered
• Redundant information
• Important information appeared too far down the page

Product goals:

• Surface entity information higher on the page
• Create clearer differentiation between profile types
• Provide easy access to integrity information
• Meet Privacy and Legal requirements
• Declutter and reorganize the header and details layout

Strategy

The initial request originated from an executive concern that the profile surface felt cluttered. This triggered a broader redesign that had implications for privacy, integrity, and policy adoption. Before moving into concepts, I collected requirements from each partner team and audited the current experience. The priority was understanding what integrity signals needed to remain visible and how far they could move without breaking compliance.

Process

We examined the four existing profile types and confirmed how visually similar they looked. There were minimal cues to help users understand whether an account represented a person, a business, or a community.

In the current design, we displayed a "Page" label in the About section. Feedback showed that the label felt heavy, consumed more space than necessary, and users did not understand what it represented — but removing it would violate privacy and integrity requirements.

Early explorations

For the first pass, we explored adding a new visual element that could signal entity type without removing required disclosures — testing variations with icons, subtle labels, or category only. Feedback from partner teams established firm constraints:

• The integrity entry point must remain visible (currently an "i" icon)
• No new elements could live near the avatar, which reserves space for online status
• Category labels were not clearly tappable
• We needed subtler visual language and should explore ways to reduce header height

Iterating on the header

Because scale and localization made new elements costly, we shifted toward structural changes. We tested relocating labels into the header or into the details area — each approach requiring review from Privacy, Legal, and Integrity.

At the same time, we pursued a smaller header as part of our content-first initiative. We tested:

• Left-aligned profile photos
• Reduced avatar size
• Reduced cover photo height
• Condensed informational elements into fewer lines

The net gain was about 54 pixels of usable space, although this raised new privacy requirements we needed to meet.

Final solution

After rounds of reviews with Legal, Privacy, and Integrity, we landed on a structure that preserved required text and entry points, reduced visual clutter, increased space for primary content, improved accessibility, and did not increase section height.

We applied this approach across Additional and Professional Profiles, and then to Community Profiles. We also accounted for existing differentiators like varying primary CTA placements, usernames for eligible entities, and the verified badge. Together these produced a clearer hierarchy between the four main profile types.

User testing insights

• Too much visible information weakens UX — progressive disclosure works best unless the signal is critical
• Regional differences matter — participants in Brazil cared strongly about understanding advertising capability
• Type clarity is valued — users wanted clarity on whether a profile was personal or professional
• Admin info was mixed — there was mixed value in knowing admin information
• Feature awareness was low — most participants did not know that different profile types had different features

What we learned

• Integrity requirements are non-negotiable — integrity information must remain within one click
• Differentiation does not need to rely on a text label alone
• The entity label and integrity entry point do not have to be the same element

Based on these learnings, we updated the design so the arrow navigates directly to integrity information, and all required integrity elements remain visible within one tap.

Project outcomes

• 64% increase in scroll engagement
• 8% reduced bounce rate
• 15% more profile and cover photo area
• All integrity and compliance requirements preserved

Additional Profile — Before and after:

Community Profile — Before and after:

Conclusion

At Facebook's scale, the most valuable design skill isn't visual — it's building alignment across teams with competing priorities. The constraints weren't the enemy of good design; they were the design problem itself. We continue monitoring performance and usage around sizing, spacing, interaction patterns, and content scanning behavior to optimize clarity across profile types.