Google Research · 2022–2023 0→1 Launch Computer Vision AI Incubator

0→1 AI Fitness App: Novel HCI & Computer Vision Research from Discovery to 100k-User Launch

Building the UXR practice from scratch for a novel computer vision fitness application incubated inside Google Research — establishing mental models, validating AI interactions, and driving two parallel product launches across web and mobile.

Computer vision pose estimation for AI fitness
Role
UX Researcher (Lead)
Timeline
Apr 2022 – Mar 2023 · ~10 months
Methods
Foundational IDIs, usability testing (remote & in-lab), literature review, heuristic evaluation, surveys, workshop facilitation
Impact
AI accuracy improved 50%→90%; AI transparency +40%; web app + mobile beta launches; 4 validated user personas; curb-cut accessibility improvements
Context

Creating multi-modal AI health applications to democratize fitness

Google Research was developing a novel AI-enabled fitness application using camera-based computer vision and 3D movement modeling — technology with no established UX precedent. The product would use a device's camera to analyze a user's movement in real time, provide form feedback, and track fitness progress without wearables or specialized equipment.

The ambition was significant: democratize access to high-quality fitness coaching by making AI the coach. But when I joined, the team had started development with almost no foundational user research — no validated user segments, no mental models for how people would relate to camera-based AI exercise feedback, and no UXR infrastructure at all.

Two parallel products were in development: a web application (further along, approaching an internal launch) and a mobile application (earlier stage). My role was to build the UXR practice from zero and support both — running at the tactical speed of sprint cycles while simultaneously doing the foundational research that would shape long-term product strategy.

The core challenge: No established mental models existed for camera-based AI fitness interactions. Users had never been coached by a computer that watched them move. Every assumption about how people would perceive, trust, and interact with this technology had to be validated — and there was no prior research to build on.

Computer vision pose estimation — AI analyzing human movement in a home environment
Computer vision pose estimation — the core AI technology mapping 3D human movement in real time across web and mobile surfaces.
The challenges

Five compounding problems at the start

The research challenges weren't just about the product — they were structural. Walking into an incubated team mid-development meant inheriting a set of compounding problems that had to be addressed in parallel:

Problem 01
No established mental models
Camera-based AI fitness coaching had no UX precedent. Users had no framework for how to interact with, trust, or evaluate an AI that watches them exercise.
Problem 02
No UXR infrastructure
Research operations didn't exist. No intake process, no tracking, no OKRs, no established relationship between research and product sprints.
Problem 03
Development started without foundational research
Both web and mobile products were already in active development. User segments, journeys, and mental models were assumed, not validated.
Problem 04
Usability never evaluated
Neither product had been through usability testing with real users. Red-flag issues were unknown, and the AI interaction model was unvalidated.
Problem 05
Two products, one timeline
Web app and mobile app were developing in parallel with separate teams and timelines. Research had to serve both without duplicating effort.
Research objectives
Five goals driving the program
Understand fitness user journeys; prioritize AI features; assess usability and accessibility; evaluate and improve AI models; support cross-functional launches.
Research program

A phased program across two parallel products

The research program ran in two phases. Phase 1 addressed the web app's immediate tactical needs — getting usability research running fast enough to support active sprints. Phase 2 ran the foundational and mobile research in parallel, with insights flowing back to inform both products.

Phase 1 · Step 1
UXR operations setup
Conducted stakeholder interviews to understand team needs. Led an "Intro to UXR" workshop. Created a research intake form, tracker, scoped timelines, and set quarterly OKRs. Built the infrastructure that would allow research to run at sprint cadence.
Phase 1 · Step 2
Usability round 1 — web app (n=10)
60-minute semi-moderated remote IDIs with 10 fitness users. No prior product feedback existed — sessions combined usability testing with behavioral probing. Surfaced P0 issues, light-weight user journeys, and — critically — insights from users with temporary vision impairment that led to "curb cut effect" accessibility improvements.
Phase 1 · Step 3
Usability round 2 — web app post-redesign (n=7)
30-minute remote sessions after P0 issues were addressed. Validated that redesigns improved the experience. Comparative analysis confirmed AI accuracy improved from 50% to ~90%, and AI transparency improved by 40% — users now understood when AI was active and what it was doing.
Phase 2 · Step 4
Literature review — user segmentation
40+ peer-reviewed journal references. Cross-referenced with Fitbit and Google Fit user data. Identified missing segmentation for users with impaired hearing, vision, and mobility — directly impacting accessibility strategy. Generated baseline data for recruitment screeners and persona development.
Phase 2 · Step 5
Foundational IDIs — end users (n=15) + fitness instructors (n=15)
Two separate 60-minute IDI studies. End user study explored fitness journeys, device ecosystems, and AI personalization expectations. Fitness instructor study mapped how instructors create content, pain points in online delivery, and where AI could augment human-in-the-loop coaching. Both produced journey maps, personas, and strategic insights for mobile feature prioritization.
Phase 2 · Step 6
Heuristic evaluation — mobile app (5 expert reviewers)
Rather than jumping straight to external user testing on the mobile app, I steered the team toward a faster, higher-resolution method: a heuristic evaluation using Jakob's Ten Usability Heuristics with 5 expert reviewers from UXR, UXD, Engineering, and Product. Planning took 3 days; testing and prioritization completed in under 48 hours. Identified 10+ P0 issues — 3× faster than external testing would have been at this stage.
Phase 2 · Step 7
Iterative mobile usability — in-lab + remote (n=25+)
2 in-person lab studies and 1 remote study after P0 issues were resolved. 30-minute sessions testing ~6 core tasks. Lab setup allowed livestreaming to team members in real time — increasing stakeholder understanding and buy-in. Produced a usability tracker for engineering to execute agile sprints directly.
Phase 2 · Step 8
Workshops — CUJs, personas, and team alignment
Led a series of cross-functional FigJam workshops with designers, engineers, PMs, and project managers to build validated user journeys, personas, and critical user journeys. Including stakeholders in building these artifacts drove team buy-in and gave the whole team a shared user model to make product decisions against.
Phase 2 · Step 9
Feedback survey redesign + validation quant
Redesigned CSAT/NPS surveys to improve completion rates while capturing AI-specific perception signals (perceived accuracy, lag, transparency). Created a Qualtrics skip-logic survey to validate qualitative device ecosystem insights with a larger sample — quantifying job success and importance ratings to drive mobile feature prioritization.
AI research framework

Defining what "good" looks like for AI interactions

One of the most substantive contributions of this project was developing the team's framework for evaluating AI interactions — a set of dimensions that gave engineers, designers, and PMs shared vocabulary for what "good" meant for a camera-based AI product in a health context.

Transparency
AI systems should be clear about what they're doing and why. Users should understand the logic behind recommendations — building trust through legibility.
Trust
Users need to feel the system is reliable, secure, and behaves as expected. Demonstrated through accurate performance and responsible data handling.
Bias & Fairness
AI systems must avoid and mitigate algorithmic bias. Particularly critical in fitness contexts where body type and movement diversity are significant variables.
Control
Users should feel they can adjust or override AI decisions. Autonomy over AI interactions is essential to sustained engagement.
Accessibility
AI tools must be inclusive for all users regardless of ability. The "curb cut effect" discovery — insights from low-vision users improving the experience for all — came directly from this lens.
Personalization
AI's capacity for personalization must be balanced with privacy. Users should feel recognized, not surveilled — a nuanced distinction surfaced in foundational research.
User personas

Four validated personas from 40+ interviews

Through cross-functional workshops, I led the creation of 4 user personas based on real user data from foundational IDIs, usability surveys, and cross-product data from Fitbit and Google Health. Personas were prioritized by opportunity and influence — the 4 most strategically relevant segments out of 5 identified.

Four user personas for the AI fitness application
Four validated user personas developed through cross-functional workshops — used across UXD, Engineering, Product, and Marketing for product decisions.
Sample persona detail — Sarah Johnson, part-time administrator, fitness app user
Sample persona: Sarah Johnson — representative of the busy parent segment, a high-opportunity user group for flexible, AI-personalized fitness.
Impact

From 0 to two product launches

50→90%
AI model accuracy improvement driven by usability research findings
+40%
AI transparency improvement — users understanding when and how AI was active
40+
Foundational interviews with end users and fitness instructors
4
Validated user personas adopted across product, design, engineering, and marketing
2
Product launches supported — web app internal launch + mobile beta
48hrs
Heuristic evaluation turnaround — 10+ P0 issues identified, 3× faster than external testing

Specific research artifacts, session recordings, and raw data are protected under NDA. Illustrative personas and diagrams shown above represent the structure and format of actual deliverables. Available for discussion in a portfolio review. Request one →

Reflection

What 0→1 AI research taught me about operating in incubators

The most important skill this project developed was knowing which altitude to operate at in each moment. Sprint-level research — fast, targeted, immediately actionable — kept the team moving. Foundational research — slower, strategic, longer-horizon — gave the product a foundation it wouldn't have had otherwise. The challenge was running both simultaneously without letting either collapse into the other.

The heuristic evaluation decision was one of the highest-impact calls I made. Rather than defaulting to external user testing when the mobile app was ready, I recognized that a well-run HE with the right experts would surface more critical issues, faster, with higher fidelity — and it did. Knowing when to use a faster, cheaper, higher-resolution method over a slower, more expensive one is a research skill that doesn't get taught enough.

Building UXR infrastructure from scratch also clarified something about embedded research: the process artifacts — the intake form, the tracker, the OKR framework — are not administrative overhead. They are the mechanism through which research becomes trusted and used. Teams that don't have them treat research as optional. Teams that do treat it as a product requirement.

Back to all work Next → Verily — B2B ML research for Living Health
Editing Click any highlighted text to edit