FitForAll: AI-Powered Fitness Coach for Underserved Communities
Solo-built AI fitness coaching app for users from African, South Asian, East Asian, and Hispanic backgrounds. Designed, built, and shipped the full MVP to validate the hypothesis: if a fitness app personalises around cultural food heritage rather than Western defaults, it achieves meaningfully better activation. Live at fitforall.up.railway.app.
Challenge
The global fitness app market exceeds $15B but was built almost entirely for one cultural context. Most apps assume users eat chicken breast and broccoli, work out at a gym, and have no cultural occasions that interrupt a routine. Through conversations with friends and community members, I kept hearing: I tried MyFitnessPal but it doesnt know jollof rice. This was not niche - it was a structural gap. The failure modes: food irrelevance (databases skew Western), no cultural accommodation (apps treat Ramadan, Diwali, family meals as cheat days), and retention collapse (users disengage within the first week).
Approach
- • Culture-first onboarding: Put culture selection as Step 2, immediately after goal selection. Signals to the user that this product was designed for them, not adapted. Created immediate relevance that buried-preference competitors miss.
- • AI as the core product: Made the AI the primary interface rather than a database-driven tracker with chat added on top. Every recommendation flows through a culturally-aware system prompt. The product improves through prompt engineering, not content curation.
- • No account creation: Users go from landing page to personalised experience in under 60 seconds. For an MVP relying on organic growth, removing authentication was correct. Trust is earned through value, not logins.
- • WhatsApp as the growth channel: Built one-tap share with pre-written copy that references the users culture. For African and South Asian users, WhatsApp is the default sharing layer. It costs nothing and fits naturally into how these communities communicate.
- • Engineered strict output format: Maximum 3 sentences for welcome messages, always end with a question to drive engagement, no generic disclaimers. The most impactful change - reducing welcome from 200+ words to 2-3 sentences - improved first-interaction engagement noticeably.
- • localStorage over database: Zero backend cost at MVP scale. Tradeoff: data lost if user clears browser. A known limitation, acceptable for a learning exercise.
Outcome
- • Onboarding completion rate reached 65% (vs 35-40% industry benchmarks) by making culture selection the first interaction - validating the core hypothesis that cultural relevance drives activation.
- • Shipped a functioning MVP to test the hypothesis with real users. The product proves that prompt engineering can encode cultural context as a product feature, not an afterthought.
- • Navigated real technical constraints: tiktoken Rust compilation failure (removed dependency), Python 3.14 incompatibility (pinned 3.11), input validation gaps (added HTML5 + JS guards). Every build failure taught me something a PRD cannot.
- • Established a repeatable framework for cultural product validation that can extend to any underrepresented market segment. The insight: users stayed because advice fit their lives, not because of feature depth.