Selected work

Examples of how I help teams

Many of the details of my work are confidential, but the patterns are consistent: clarifying decisions, grounding AI in reality, and aligning teams.

These projects show what that looks like in practice across AI products, complex operations, and subscription businesses.

31% ↑ usability · 28% ↓ time-on-task · New roles created · Industry impact

Scaling human–automation teaming for stratospheric internet

Google X · Complex operations · Human–automation teaming

Challenge. Loon needed to scale from ~60 to hundreds of stratospheric “floating cell towers” without scaling flight-operations headcount linearly. The internal model treated operators like pilots — a mismatch for how the system actually worked.

What I did. Led research across workflow analysis, expert interviews, policy reviews, and usability testing, revealing a core category mistake in how the system was conceptualized.

Reframe. Showed that Loon functioned as airborne infrastructure, not aviation — and that operators were acting as automation supervisors (similar to SREs), not pilots.

Impact.

  • 31% increase in system usability (SUS 65 → 85.3)
  • ~28% reduction in exception-handling time-on-task
  • Two new operational roles created to support scaled supervision
  • Reframe adopted across leadership, legal, policy, and engineering
  • Influenced NASA ConOps, ICAO, ASTM, and HAPS Alliance
  • Enabled a sublinear scaling strategy for global operations

Further reading. Published case study (Journal of Air Traffic Control, PDF)

↓ confusion & misinterpretation · ↑ trust & clarity · AI/UI patterns

Reducing AI confusion and risk in Search

Google Search · AI personality & trust · Product strategy

Challenge. As LLMs entered Search (AI-Mode, AI Overviews, Gemini), chat-style patterns led users to over-anthropomorphize the system, creating unrealistic expectations and trust issues.

What I did. Broke down “personality” into actionable dimensions, designed and shipped a multi-response prototype for AI-augmented search, and created metrics to evaluate perception, trust, and expectation alignment.

Impact.

  • Reduced anthropomorphization and user confusion
  • Improved trust and expectation alignment with model capabilities
  • Influenced AI-Mode design direction and upcoming multi-response interaction patterns
  • Framework adopted across Search & DeepMind
  • Secured cross-PA funding and exec sponsorship
24-hr turnaround · Drivers clarified · Pricing fixed · Retention improved

Reversing a subscription churn spike

Native Instruments · Subscription strategy · Rapid mixed-methods

Challenge. Sounds.com saw a sudden churn spike. Leadership needed clear answers and a plan within 24 hours.

What I did. Analyzed 597 open-text exit-survey responses using hybrid manual + NLTK analysis, combined with stakeholder interviews, to pinpoint high-confidence churn drivers.

Impact.

  • Reframed churn into specific, addressable drivers
  • Informed new pricing & packaging (subscriptions + credit packs)
  • Enabled targeted win-back campaigns
  • Shifted roadmap toward quality, IA, and core usability
  • Secured investment in ongoing product-health research

Interested in something similar?

If you’re facing related questions in your own AI work, we can talk about how these patterns map to your context — from strategy and research to concrete product and system decisions.

Talk about a project