When we started NUA almost two years ago, generative AI was not yet viable for production use. But even then, we had a clear intuition: we didn't want to build another training app — we wanted something closer to a coach. Combining human-like presence with algorithmic rigor.
What felt fundamentally broken
As athletes ourselves, two things felt broken in every tool we used.
First, opacity. Decisions arrived without explanation — "trust the plan." Trust erodes over weeks until you abandon the plan entirely. Second, isolation. You get a spreadsheet, not a companion. The plan doesn't know you're tired, stressed, or doubting yourself.
A truly intelligent coach must solve both: make real decisions and make you feel accompanied.
Our first approach — and where it broke
Our MVP focused on what we thought mattered most: a strong training engine aimed at serious athletes, a conversational interface via WhatsApp to emulate the warmth of a real coach, and a conscious decision not to build dashboards — tools like Strava, Garmin, or Intervals.icu already do that well.
The idea was simple: training logic in the engine, human experience in the chat.
That worked… until it didn't.
Once something talks like a coach, athletes expect it to reason like one — across any situation. But our conversational layer was explaining decisions it didn't actually make. When reality drifted outside predefined paths, the cracks showed.
You cannot separate the entity that decides from the one that explains and expect trust to hold.
Coaches are closed-loop systems
What we eventually realized is that a coach is not a set of features. It's a closed loop: understand initial conditions, plan, observe execution, explain feedback, re-plan when reality diverges.
And this loop exists at multiple layers simultaneously — season goals, macrocycles, weekly structure, day-to-day readiness.
If different parts of the system reason independently at each layer, coherence is lost. Your Tuesday intervals exist because of your March race, your fatigue from Sunday's ride, and the fact that you slept poorly last night.
Where generative AI actually helped
When generative AI became reliable enough, we initially thought of it as a UX upgrade: more empathy, more flexibility.
But the deeper value was elsewhere. Not removing rules — abstracting them.
Instead of thousands of if/else branches, we now define what strategies are allowed, what constraints must never be violated, and what data matters at each decision layer. Within that controlled space, AI selects and justifies decisions — but never invents capabilities the system doesn't have.
Today NUA runs as a hybrid system: algorithmic structure guarantees safety and consistency, while AI operates as a constrained decision layer with 60+ specialized agents, governed by sport science rules. Crucially, the same entity both decides and explains.
Absorbing complexity, not exporting it
Internally, we track a large number of metrics — session execution, load dynamics, readiness trends, performance signals.
But the athlete should experience clarity, not cognitive load.
Dashboards, TSS charts, CTL curves — these externalize complexity onto the athlete. They make you the analyst of your own performance. Our job is to absorb that complexity, not export it. To tell you what to do today, and why.
Athletes don't want to allocate mental energy to micro-decisions a system can reason about more consistently. They want to focus on execution, sensation, and intent. Good coaching has always done this. AI simply allows it to scale.
Still a work in progress
We don't consider NUA finished — far from it. Each iteration comes from realizing where our previous mental model of a coach was incomplete.
But if we've learned one thing, it's this: building an AI coach is less about adding intelligence, and more about preserving coherence, trust, and explainability as complexity grows.