Product
Training loop
Set a goal, follow a structured plan, and adapt sessions based on real fatigue. Workouts are designed from elite coaching programs defined with Maxime Ducret.
Side build · Rowing app
Splash
A rowing app built entirely from scratch in ~25 hours using AI assistance. Test-driving complete product execution from ideation to working cross-platform prototype.
Product
& full-stack
2026
Overview
Coaching logic turned into code
Team & context
AI-assisted sprint
A side-project built to test AI development tools (Cursor). Scoped to explore rapid product and technical execution from zero to prototype.
My Role
Full ownership
Product strategy, UI/UX design, domain logic extraction with a coach, and full-stack implementation using Cursor, Supabase, and Vercel.
The opportunity
Filling the void in rowing software
The friction
No guidance
Runners and cyclists have dozens of training apps. Rowers don't. They repeat the same generic sessions without knowing if they're on track for their goals.
The challenge
Software that acts like a coach
Deliver quality coaching guidance without pretending AI can replace a human. Turn real-world coaching principles into a structured but flexible digital plan.
How it works
A training engine guided by fatigue
Set a goal, follow a structured plan, and let your actual state of recovery dictate the next optimal workout.
Intention
Set your goal
Choose a distance (500m to 6k), enter your current time, and a target date. Splash calculates a realistic improvement target based on training science.
Programming
Get your plan
The app generates workouts using Maxime's coaching programs. Each session is uniquely mathematically adapted to your current level and goal.
Execution
Train and adapt
Log sessions (rowing or other sports). The app tracks fatigue and adjusts the plan week-to-week to fit how you're actually feeling, like a real coach.
Execution
Iterating on code, not wireframes
Product strategy
Cut ruthlessly
Stripped the scope down to one core flow: set a goal, get a plan, adapt to fatigue. Just the athlete's training loop, enabling an end-to-end build in under 30 hours.
Domain modelling
Encode expert knowledge
Worked with a champion turned full-time coach to translate his training principles (intensity zones, progression, race prep) into usable algorithms.
Training experience
Build around fatigue
Rather than tracking volume blindly, the app monitors fatigue across all activities to recommend what to do next. Plans adapt to life, not just theory.
Execution layer
Ship with AI assistance
Used Cursor with Supabase/Vercel. Prototypes in code instead of Figma. Generated, refined, and enforced architectural standards through strict AI prompting.
Result
From idea to working prototype in 25 hours
A testable product
Not a mockup
Rowers can create goals, receive workouts, log sessions, and watch fatigue evolve. It's an early-stage but fundamentally complete beta.
Proof of concept for AI dev
New workflows
Demonstrated that AI excels at rapid prototyping if someone owns architecture. Product thinking is the bottleneck now, not boilerplate.
Unplanned traction · one month after release
300+ users
Zero marketing effort, just word of mouth among rowers after a few social posts.
1,000+ sessions
Training sessions logged by real rowers, proving the core loop works.
Positive retention
Users keep coming back without prompts, the product creates a weekly habit around structured training.
Retrospective
What I learned
Speed favours feedback over speculation
Getting to a testable app in twenty hours changes the equation. You can afford to learn from users much earlier rather than over-engineering on paper or trying to future-proof everything before launch.
Scope discipline matters more, not less
Faster building means it's easier to overbuild. Saying no to everything beyond the core training loop was essential to keeping this shippable.
AI is a fast junior, not a senior
Cursor drafts code incredibly quickly, but it doesn't replace architecture. It builds complex tangles if left unattended. Visual hierarchy, interaction details, and UI polish still require strong human product judgment.