Selected work / 01
EyeXP
An AI coach that watches your technique. Before launch, I set out to watch the product the same way.
- Role
- Product ManagerMetrics, analytics, research and growth experiments — next to the product design
- Product
- AI technique coachPose analysis for boxing and fitness drills
- Stage
- Closed alphaPre-launch, iOS
- Stack
- PostHogAnalytics, session replays, surveys, feature flags
- Focus
- Activation · RetentionAnd one growth experiment before launch
Measure now, while it is cheap.
Before the alpha, I defined what we would need to know at launch — and made the case for building it in now.
How I put it to them
I put it to the founders as their gain, not my process: before launch we need to see where people get lost and what brings them back. Events added after launch mean rework — and the first weeks, the most telling ones, are never recorded at all.
- AlphaA few days of one engineer
- BetaRework in screens already shipped
- LaunchThe first cohorts, never recorded
- ScaleDecisions made on guesses
Three things we agreed
- 01
Who builds it
One mobile engineer, a fixed budget of hours, events shipped with the alpha build.
- 02
Which tool
Chosen on criteria, not habit — see below.
- 03
How we use it
A weekly review with the founders: the funnel, the cohorts — and one decision taken from them.
One number to steer by.
One number the whole team steers by, four that move it, and two we never trade for growth.
North Star
Completed workouts per active user, per week
Activation
First workout done within 72 hRetention
Back on day 1, 7, 30Frequency
Workouts per weekQuality
Cues applied on the next attempt
Guardrails
- Analysis failures per workout
- Time from upload to feedback
Activation
First workout completed within 72 hours of sign-up.Why this definition
The moment someone has felt the product work on their own body. Seventy-two hours covers a real training day, while the intent is still fresh.
Retention
Share of people back on day 1, 7 and 30 — and by weekly cohorts.Why this definition
Day 1 tells whether the first session landed; day 7, whether a habit started; day 30, whether there is a product.
Onboarding funnel
Downloaded → signed up → onboarding → plan → first workout started → finished.Why this definition
Six steps, each one an event, so every leak has an address.
- 01Downloaded
- 02Signed up
- 03Onboarding done
- 04Plan ready
- 05First workout started
- 06First workout finished
Not chosen as the North Star: downloads (they measure marketing), sessions (opening the app is not training), the precision score (someone can score well and never come back).
Every event named before it is built.
One table the engineers could build from, and one rule they could remember.
object_action — lower case, underscores, past tense.
signup_completedAn account is createdmethodFunnelonboarding_step_viewedEach onboarding screen opensstepstep_nameFunnelonboarding_completedThe last step is doneduration_sskipped_stepsFunnelplan_generatedThe first plan is readygoallevelsessions_per_weekFunnelworkout_startedThe first rep is recordedworkout_typesourceActivationworkout_completedEvery set is finishedworkout_typeduration_ssets_doneNorth Staranalysis_completedFeedback is readyprecisioncues_countprocessing_mscamera_stableGuardrailfeedback_viewedA result is openedcue_identryQualityai_coach_message_sentA question goes to the coachtopiclength_bucketEngagementfeedback_submittedAn in-app survey is answeredsurvey_idratingResearchresult_sharedA result card is exporteddestinationGrowthOne tool, chosen on purpose.
The questions of the alpha and the experiments of the launch should live in the same data.
Decision
PostHog: analytics, replays, surveys and flags in one place — what we learn in the alpha and what we test after launch, in one dataset.
Had the app already shipped Firebase, I would have kept it: one SDK fewer is worth more than one feature more.
I walked every path myself.
Analytics break quietly. Before the first tester arrived, I installed the alpha and went through every route, watching each event land.

- Fires onceNo duplicates on re-render, background or retry.
- Properties completeEvery property present, with the right type and values.
- One person, one IDThe anonymous visitor and the signed-up user merge into one.
- Time is rightTimestamps and time zones survive offline sessions.
- Nothing personalNo video, text or body data in any payload.
Why this matters
This is the dull step, and the one where most tracking plans fail: a duplicated event inflates a funnel, a missing property hides a segment, a broken identity merge turns one person into two. Each would have poisoned every chart that followed.
Two views, read every week.
The onboarding funnel up to the first finished workout, and retention by weekly cohort.
- Downloaded64
- Signed up5789%
- Onboarding done3154%
- Plan ready3097%
- First workout started2480%
- First workout finished1979%
07Numbers & voices
The data and the people said the same thing.
Seven alpha testers, 20–30 minutes each: what was unclear, why they stopped, what they liked.
The interview guide
- Walk me through the first time you opened EyeXP.
- Where did you hesitate?
- When did you stop — and what were you doing then?
- What did you expect to get back?
- What would make you open it tomorrow?
- 57Step 1
- 54Step 2
- 52Step 3
- 33Step 4
- 31Step 5
−37% after step 3
“I hadn’t thrown a single punch yet, and it wanted me to film my whole body.”
Tester, 29“I thought it would record me all the time. That felt like a lot.”
Tester, 34“Propping the phone up in my kitchen — I closed it and said I’d do it at the gym.”
Tester, 24“I didn’t know what I’d get back for it.”
Tester, 41“Once I saw the cue on a jab, I got it. I just saw it too late.”
Tester, 31“The one tip with a timestamp is the best part.”
Tester, 27“I wanted a plan for the week, not just one drill.”
Tester, 36
In the data, about four in ten testers left at step 3. In the interviews, five of seven said they didn’t see why the app needed the camera before it had shown them anything.
Show the value, then ask.
Step three asked people to prop up the phone and film their whole body before they had seen anything in return.
Next cohort, same build otherwise
Onboarding completed
+24 pts
Activated within 72 h
+14 pts
DirectionalSmall samples: the direction is clear, the size is not. The next cohort checks it again — with a feature flag, once there is traffic.
What stayed the same
Kept the same, to compare like with like: the invite channel, the rest of the build, the tester profile.
Test the promise before the product.
Two landing pages, one waitlist form, two promises. Meta ads with the same audience and budget on each.
Variant A
“An AI coach that adapts to you.”
- Ad CTR
- 1.1%
- Landing → waitlist
- 18%
- Cost per lead
- €3.90
Variant B
“20-minute workouts. No gym.”
- Ad CTR
- 1.6%
- Landing → waitlist
- 24%
- Cost per lead
- €2.10
- Budget€150–300 per variant
- AudienceThe same, split evenly
- ReadAd CTR · landing → waitlist · cost per lead
The concrete promise won on every measure: people respond to time and place before they respond to technology. The AI becomes the reason it works — not the headline.
01 Review
Check the take before it counts.
A shaky camera is named calmly, with a retake — bad input never becomes bad feedback.
attempt_recordedcamera_stable: falseattempt_retaken
02 Analysis
Work you can watch.
Frames, keypoints, comparison — progress you can trust, and permission to leave.
analysis_startedanalysis_completedprocessing_ms
03 Feedback
One cue first.
The most important correction on top, pinned to the moment in the clip; the rest wait their turn.
feedback_viewedcue_idworkout_completed
04 Share
A result worth posting.
Progress turned into a card — and a growth loop we can count.
result_shareddestination

A habit of measuring, before there was anything to lose.
- Defined the North Star, activation and retention for a pre-launch AI fitness app
- Wrote the tracking plan: eleven events, their properties and one naming rule
- Chose the tool on criteria and verified every event by hand
- Built the funnel and cohort dashboards the team reads weekly
- Found the onboarding leak where the numbers and seven interviews agreed
- Fixed it and compared the next cohort — honestly, as a direction
- Designed a value-proposition smoke test to run before launch
Next
- Run the onboarding change as a flagged A/B once traffic allows.
- Link cue quality to retention: do people who apply cues come back more?
- Add a guardrail alert when analysis failures rise.


