Case Study · Consumer App · Health & Fitness

How We Built an AI-Native Health & Fitness App
That Adapts to Every User in Real Time

A health and fitness brand wanted a consumer app where AI wasn’t a marketing line — it was the entire experience. We built the recommender, the planner, and the adaptive feedback loop in one integrated product.

100% AI-driven user experience
— no static plans
6 Weeks Kickoff to v1 launch
on iOS & Android
Real-time Personalisation across nutrition,
training, and recovery

A health and fitness brand approached us with a clear vision: they didn’t want another generic workout app with a static plan stapled to onboarding. They wanted an app where AI ran the entire experience — adapting in real time to what the user did, what they ate, how they recovered, and how their goals shifted.

The market was full of apps with “AI” in the marketing copy but a decision tree in the code. They wanted the opposite: an experience where every recommendation, every plan adjustment, every nudge came from a real understanding of the individual user, not a bucketed segment.

Their constraint was the standard consumer-app one: ship something real, fast, that early users would actually open every day. No 9-month build. No discovery phase. Live in weeks.


The first agency they spoke to pitched a 6-month “AI-first” build that, when we reviewed the proposal, was actually a rules-engine wrapped around a marketing claim. The second pitched a no-code MVP that wouldn’t have scaled past 1,000 users.

We came in with a different proposition: we’d architect the app as AI-native from week 1, with the recommender, the planner, and the feedback loop built as core systems, not as features bolted onto a CMS. And we’d ship in 6 weeks.


The AI-Native Consumer App

Four core systems, all working together in real time.

Module 01 · Onboarding

Onboarding & Profile Inference

A conversational onboarding flow that goes beyond a form. The app asks open-ended questions, parses free-text responses, and builds a rich user profile across goals, current fitness baseline, dietary preferences, schedule constraints, and motivational style.

Module 02 · Recommender

The AI Recommender

A personalised recommender that generates daily workout, nutrition, and recovery plans for each user, conditioned on their profile and their last 7–14 days of behaviour. Not a rotating library of pre-made plans — actual generative output tuned to the individual.

Module 03 · Planner

The Adaptive Planner

The plan adapts continuously. If the user skipped yesterday’s workout, today’s recommendation accounts for it. If they logged a heavier meal, the next day’s nutrition target shifts. If they reported soreness, recovery gets prioritised. The planner isn’t running a rules tree — it’s running an LLM-based reasoning layer over the user’s state.

Module 04 · Feedback

The Feedback & Nudge Loop

The app learns from every user action — what they completed, what they skipped, what they liked, what they swapped. That signal flows back into the recommender so the next day’s plan is genuinely smarter than the last.


Frontend
React Native
iOS & Android from one codebase
TypeScript
Type-safe component layer
Tailwind
Design system & tokens
Backend
Node.js
Core API gateway
PostgreSQL
User profiles & session history
Redis
User state cache
AI Layer
Claude (primary)
Reasoning & plan generation
OpenAI (fallback)
Redundancy layer
Embeddings + vector store
Behavioural retrieval
Recommender
LLM reasoning layer
Over structured user state
Similar-user retrieval
Hybrid personalisation
S3
Media & asset storage
Infrastructure
AWS (EKS + RDS + S3)
Compute, database, storage
CloudFront
Global CDN
Firebase
Push notification infrastructure
Analytics
Event instrumentation
Every user action captured
User-state warehouse
Feeds the recommender directly
A/B testing infra
Production experimentation layer

Six sprints, each with a hard gate. No slide decks, no “it’ll be ready next week.” A working build every week, regardless of what was unfinished behind the scenes.

Week 1

Scope, AI Architecture & Onboarding Wireframes

Locked the architecture: four modules, one data model, one AI layer. Defined the user profile schema, the behavioural signal format, and the inference pipeline structure. Set up the React Native build pipeline, the AWS environment, and the CI/CD flow. Produced high-fidelity onboarding wireframes by Thursday. Founder reviewed on Friday. We had agreement on every screen before a single line of production code was written.

Week 2

Onboarding Shipped

The full conversational onboarding flow was live and connected to the backend by Friday of week two. Free-text goal parsing, fitness baseline inference, dietary preference capture, schedule constraints, motivational style. The output: a structured user profile that the recommender could act on immediately. The founder completed onboarding on their own device on Thursday. The profile it generated was accurate enough that they asked us to add it to the demo script.

Week 3

Recommender v1

The AI recommender was live by end of week three. Complete onboarding, receive a fully personalised plan across training, nutrition, and recovery. Not a templated plan with your name on it — actual generative output conditioned on your profile, reasoned through the LLM layer, and rendered into a usable daily structure. The Friday demo was the first time the founder saw the system produce a plan that surprised them. That was the moment the product became real to them.

Week 4

Adaptive Planner & Feedback Loop

The planner connected to the recommender in week four. Complete a session: the remaining week updates. Log a meal: tomorrow’s nutrition target shifts. Report soreness: recovery gets prioritised. The feedback loop closed: every user action became a signal that flowed back into the next recommendation. The Friday demo was a full end-to-end experience — onboard, get a plan, complete a workout, see the plan adapt. The AI’s reasoning was visible in the UI, not hidden behind a seamless facade.

Week 5

Polish, A/B Infrastructure & Push Notifications

Performance pass: 60fps on all animated transitions, recommendation latency under 400ms P95, cold-start under two seconds on a median device. A/B testing infrastructure deployed — production experimentation ready from day one of live. Push notification system live via Firebase: workout reminders, streak nudges, adaptive check-in prompts. Beta build shipped to 40 users via TestFlight and Android internal testing. Seven issues found; six resolved by Thursday; one deferred with founder sign-off.

Week 6

App Store & Play Store Launch

iOS submitted Monday. Android submitted Tuesday. App Store approved in 36 hours — we had cleared every potential rejection reason in the submission preparation. Google Play approved Wednesday. Influencer launch window opened Friday with a live, publicly downloadable product in both stores. First 500 downloads in the first four hours. The product the founder had been told would take six months was live in six weeks, and it worked exactly as specified.


Four numbers that define what this app is. These aren’t growth metrics or engagement statistics — they are product architecture decisions that we committed to in week one and delivered at launch.

100%
AI-driven user experience — no static plans. Every recommendation the app surfaces — every workout, every nutrition target, every recovery suggestion — is generated by the AI layer, conditioned on that specific user’s profile and behaviour history. There is no static content library. There are no pre-written plans. If the AI goes down, the app has nothing to show. That was a deliberate architecture choice, not an oversight.
6 Weeks
Kickoff to launch on both stores. Six weeks from the first architecture session to a publicly downloadable product on the App Store and Google Play. The influencer launch window was pre-committed before we started building. We hit it. The product the founder had been quoted six months to build was live in six weeks, and it was more sophisticated than the six-month proposal.
Real-time
Personalisation materially changes over time. The plan is not personalised once at onboarding and then left to run. It updates continuously — after every completed session, every logged meal, every skipped workout, every check-in. A user’s week-four plan looks materially different from their week-one plan, conditioned on what they actually did in weeks one through three. That is what real-time personalisation means in practice.
No CMS
No static content libraries. Most fitness apps are content businesses with an AI wrapper. This one is an AI business with a content delivery layer. There is no editorial team producing workout plans. There is no nutritionist writing meal templates. The AI generates the content, conditioned on the user. This is what AI-native means as an architecture principle — not AI as a feature on top of a content product.

Every agency we talked to before Solprime had a slide that said ‘AI-powered’ and a build plan that was a rules engine in disguise. Solprime showed us the actual architecture before we signed anything. We could see exactly where the AI lived and what it was doing. That transparency was the reason we chose them, and the product they delivered proved they weren’t just saying it.

Founder, Health & Fitness Brand reference available on request

If you’re building a consumer app where AI is supposed to be the experience — not a chatbot pinned to the corner of a generic product — the question to ask any potential build partner is simple: “show me where the AI lives in your reference architecture.” If the answer is a service that gets called by the rest of the app, AI is a feature. If the answer is a reasoning layer that the whole product is shaped around, it’s AI-native.

We built this in 6 weeks because we architected the recommender, the planner, and the feedback loop as the core of the product from week 1. The app is the AI. Everything else is glue.


Building a consumer app where
AI is the experience?

30 minutes. We’ll map the AI architecture, show you what AI-native actually means in a reference architecture, and tell you what six weeks would realistically produce.