Xavier Hills

Okta needed to grow the breadth and depth of integrations independent software vendors (ISVs) build for the OIN, but lacked a clear picture of how ISVs decide what to build, what blocks them, and what would make Okta a strategic partner rather than just a platform. The stakes spanned product roadmap, developer experience investment, and the go-to-market case for emerging capabilities like Cross-App Access.

I designed a sequential mixed-methods program so each phase sharpened the next:

• Phase 1: Developer journey (qualitative). 9 in-depth interviews across software, medtech, and fintech, spanning ISVs who had built for Okta and those who build for competing platforms. I mapped the end-to-end developer journey from awareness through build, submission, and depth.
• Phase 2: Decision-maker survey (quantitative). A blind survey of 30 ISV decision-makers (VPs, directors, C-level, PMs), analyzed with cross-tabulation and chi-squared tests, reporting only statistically significant relationships.

• Customer demand drives every build decision. ISVs prioritize integrations to protect existing revenue first. A request from a large existing customer ranked as the #1 roadmap-prioritization factor, well above security improvements or competitor parity.
• The developer has more influence than Okta believed. Leadership surfaces the need, but developers decide complexity and steer toward whichever platform is easiest to build for. Survey respondents independently weighted technical complexity heavily for identity integrations, validating the interview finding.
• Submission friction caps depth. A lengthy, ambiguous validation process after submission was the single biggest deterrent to building deeper integrations beyond basic SSO/SAML.
• Cross-App Access is a latent business story. Decision-makers saw it as both cost-saving and revenue-generating (most expecting 6–20% on each), and tied it to the urgent need to secure AI agents.

The program reframed Okta's ISV strategy around customer-driven value rather than technical messaging, repositioned Cross-App Access as a business and AI-security enabler, and prioritized fixing the submission experience to unlock integration depth. Design ran a findings workshop, and the work directly scoped a downstream study on using GenAI to remove development barriers.

Log Investigator, an LLM-powered tool that turns plain English into log queries, had been announced at Oktane but never validated with real customers. Leadership needed to know whether its capabilities matched user expectations, where the usability and utility gaps were, and whether to expand access, hold, or invest more before GA.

I designed a limited early-access (LEA) study with deliberately minimal instructions to avoid biasing how participants explored the tool, pairing in-product feedback and an exit survey with 30-minute customer interviews. Success was framed against five categories: ease of use, perceived utility, intent to keep using, satisfaction, and query quality.

• Value depended entirely on experience level, and surfaced two personas. Less-experienced admins found real value (query generation removed the barrier to getting started); experienced admins expected ChatGPT-like conversational depth and felt the tool fell short. The target audience was genuinely unclear.
• Expectations were set externally. The Oktane announcement and familiarity with other LLMs created an expectation of conversational, follow-up-capable interaction that a one-prompt/one-response tool couldn't meet.
• Summarization needed depth, not just retrieval. Users wanted explanations tied to specific policies and the events leading up to a log (connecting multiple events to understand a root cause), not just faster log-finding.

I recommended defining a target persona aligned to current capabilities, adding educational content to set correct expectations, exploring conversational interaction so users could dig into issues without rewriting queries, and extending the LEA over several months to observe real adoption. The study gave the team a clear-eyed read on GA readiness and a prioritized roadmap of capability gaps; Design ran a follow-on workshop to act on the findings.

Despite rising ad revenue, Spotify's ads growth fell short of goals. ~80% of income came from direct sales, making advertiser acquisition, onboarding, and support expensive. Enterprise onboarding took 2–3 months and campaign setup often needed engineering. The team needed to understand what drives an advertiser's decision to start, and to stay.

• Literature review to establish what was known and expose the research gap on awareness and retention.
• Focus groups with the Sales org across four markets (US, EU, APAC, LATAM) to surface advertiser segments.
• Competitive analysis of Meta, Google, iHeartRadio, X, and Pinterest ad platforms to benchmark onboarding and education.
• Exploratory data analysis with Data Science to build behavioral segments and identify churn signals.
• 55 hour-long in-depth interviews with enterprise advertisers unfamiliar with Spotify, to capture the non-user perspective.
• Sales enablement survey to locate training gaps affecting conversion.

• A statistically significant link between Sales' ability to explain features and advertiser spend, pinpointing a training gap, not just a product gap.
• Behavioral segments predicting retention vs. churn, operationalized into an internal tool and a Tableau dashboard that flagged at-risk advertisers for proactive outreach.
• An interactive advertiser journey map with risk points that gave Product, Sales, and Marketing a single shared reference.

A stakeholder workshop converted findings into commitments: Sales invested in more training, Marketing built awareness content, and Product teams adjusted 2023–2024 roadmaps to close experience gaps. The work produced lasting infrastructure (segmentation, a churn dashboard, and a journey map) that outlived the study itself.

Affirm faced rising negative App Store reviews, many from customers saying they wouldn't use Affirm again. The product team wanted to improve perception of the support experience; my longer-term research goal was to address the underlying inquiry drivers inside the product. The issue was severe enough to reach the CEO's inbox.

Over a roughly three-month timeline I researched the drivers of support frustration, surfacing the emotional cost of a support experience that intentionally hid contact paths to deflect inquiries. I partnered closely with Operations, whose staffing realities shaped what was feasible, and helped shape and validate a new guided help concept.

The redesign also drove an 8% decrease in interactions needed for resolution and helped Affirm meet contractual servicing obligations for customers previously unable to reach support. A retention analysis found higher repeat-customer rates among those who engaged with support.

To improve product awareness and adoption within a navigation system limited to ~5 surfaced products, I ran a global survey to quantify awareness, followed by international in-home field research that uncovered deeper causes (unclear iconography, missing signage, confusing cross-surface navigation). I validated changes with recurring relevance surveys against usage data, and ran a 6-month diary study to learn how long navigation changes take to be noticed and which education channels (interstitials, push, in-app, badges) worked. Follow-up in-lab interviews and an A/B test partnership with Data Science shaped a more flexible, customizable navigation, covered by Adweek, TechCrunch, CNET, and others.

To deter the spread of fact-checked false information, I researched where users encountered it and how to warn them without degrading the experience. Qualitative research showed a warning alone wasn't enough: users wanted sources and the choice to view content, leading to a redesigned treatment with fact-checker context and a "See Why / See Post" pattern. I ran an eye-tracking study confirming key information sat where users' gaze naturally fell, plus international research across two countries. Follow-up work uncovered that users wanted to warn others, reshaping the sharing flow. Covered by The Verge, Mashable, and Engadget.

After any substantial shoot, photographers face the same bottleneck: manually reviewing hundreds of RAW files to find the keepers before any creative work can begin. Existing tools (Lightroom, Capture One) require human eyes on every frame. I wanted to automate that first pass so the camera roll arrives pre-sorted.

Exposé scans a folder of RAW photos (Sony ARW and others), converts each image locally using macOS's built-in sips tool, and sends a resized preview to Claude's vision API for analysis. Claude rates each shot on sharpness, exposure, composition, and subject quality, then writes those ratings, star scores and color labels, directly back into the image's XMP metadata. The ratings appear immediately in Lightroom or Capture One with no extra steps. A persistent cache means re-analyzing a large folder skips images that haven't changed, keeping re-runs fast.

Partly to solve a real personal pain point as a hobbyist photographer, and partly as a deliberate learning project, shipping a full desktop app end-to-end, from AI integration and native file I/O to packaging a signed macOS DMG. It's the kind of side project that makes the human factors of AI-assisted workflows concrete: where does automation earn trust, and where does the photographer still need to be in the loop?