Introduction
FastAI Health Coach is EXAFABS' first health and wellness app — an AI-powered intermittent fasting companion built with React Native (Expo) and Claude AI. Rather than a generic timer app, FastAI combines metabolic zone tracking, a coach that genuinely remembers your habits across weeks, multi-angle meal photo analysis, machine learning weight predictions, and a Daily Insight that learns what matters to you.
FastAI Health Coach is LIVE on the iOS App Store (v1.0 shipped 2026-05-10; v2.12.2 in review with widgets, 100-year calendar, and Apple Health). Android remains in Closed Testing on Google Play with production access applied. Just launched on Product Hunt on 2026-05-12. This post explores the architecture, how Claude powers truly personalized coaching, and how the team built — and hardened — a production AI health app over more than a hundred releases.
The Problem: Why Intermittent Fasting is Hard to Sustain
Intermittent fasting is a well-studied dietary approach for weight management and metabolic health (see, for example, de Cabo & Mattson, NEJM 2019). But for most people, it's difficult to maintain. Here's why:
- Generic Timers Aren't Enough: Standard fasting apps simply count down hours. They offer no guidance on when to eat, what to eat, or how your body is responding.
- No Personalized Coaching: Every person's body, lifestyle, and goals are different. A one-size-fits-all fasting schedule leads to burnout and failure.
- Tedious Manual Tracking: Logging meals by searching food databases is friction-heavy and reduces adherence.
- No Progress Visibility: Users weigh themselves and hope for the best, with no data-driven understanding of their trajectory or what to adjust.
- Missing Context: Apps don't connect the dots: your current metabolic state, your recent meals, your stress, your sleep—all factors that affect fasting success.
Our AI-First Solution: Every Feature Powered by AI
FastAI flips the script by building every core feature around AI. Instead of a timer with bolted-on features, we designed the app's architecture so that Claude AI, machine learning, and computer vision are the foundation, not afterthoughts.
Smart Fasting Timer with Metabolic Zone Tracking
The timer displays your current metabolic state across four phases. As you fast, your body transitions through distinct metabolic zones, each with unique physiological effects:
- Fed State (0–2 hours): Your body is processing the meal you just ate. Insulin is elevated, digestion is active.
- Fat Burning (2–8 hours): Glycogen reserves are depleting. Your body begins mobilizing fat for energy.
- Ketosis (8–16 hours): Your liver produces ketones from fat, and they become your brain's primary fuel source. This is the metabolic "sweet spot" for many fasters.
- Autophagy (16+ hours): Cellular renewal accelerates. Your body begins clearing damaged cells and regenerating cellular components.
FastAI displays which zone you're currently in with real-time visual feedback, helping you understand the science of your fast and stay motivated to reach deeper phases.
AI Coach: Claude-Powered Personalized Conversations
The core differentiator is the AI Coach—a conversational assistant powered by Claude API from Anthropic. Unlike a chatbot that regurgitates generic tips, Claude receives your full fasting context and provides truly personalized advice.
When you talk to FastAI's Coach, Claude has access to: your complete fasting history, your weight trend over time, all logged meals and their macronutrients, your current metabolic zone, your stated goals (e.g., "lose 10 pounds in 8 weeks"), and any health conditions you've shared. This context transforms the conversation from generic ("drink more water") to genuinely personalized ("you've been breaking your fast with high-carb foods on Tuesdays; try having protein and fat first to stabilize blood sugar and reduce afternoon cravings").
The Claude integration uses Convex serverless backend for reliable API calls. Conversations are streamed in real-time, and the app gracefully degrades if the network is slow or offline (falling back to cached suggestions or a simpler Q&A mode).
Coach that Remembers: Long-Term Memory Across Sessions
Most AI chatbots forget you between sessions. FastAI's Coach doesn't. We built a memory pipeline using Voyage 3.5-lite embeddings and Convex's vector search so that what you tell the Coach in week one informs what it says in week six. Tell it once that you train fasted on Tuesdays and avoid dairy, and that context flows into every future conversation — meal suggestions, fasting window advice, milestone celebrations — without you having to repeat yourself.
A nightly Convex cron extracts the most useful long-term signals from your conversation and meal history, embeds them, and stores them as searchable memory. When you next open the Coach, the relevant memories are pulled in alongside your live data — the same retrieval-augmented-generation pattern used by enterprise AI, but tuned for personal health instead of corporate documents. Want to read the engineering story? We wrote it up here.
Daily Insight: A Personalized Score That Knows Your Goal
Every morning, FastAI generates a Daily Insight — a short personalized read on yesterday's fasting, meals, and trajectory. It's not a generic "great job!" — the AI references your specific goal (e.g. "lose 10 lb by July"), your stated constraints (e.g. "no dairy, vegetarian"), and any health conditions you've shared. Hit a 16-hour fast on a high-protein day? It tells you why that particular combination accelerates your timeline. Drift off-plan two days in a row? It flags the pattern before it becomes a habit.
Multi-View Meal Photo Analysis: Better Estimates from Two Angles
Manually logging meals kills motivation. FastAI solves this with AI-powered meal photo analysis: snap a picture of your food, and the app automatically identifies the dish, estimates portion size, and calculates calories, macronutrients, and micronutrients. New in v2.9: an optional second photo from a different angle dramatically improves portion accuracy. Cross-view reasoning lets Claude triangulate volume better than a single overhead shot — particularly useful for layered dishes (curry over rice, a stuffed sandwich, a topped salad) where one angle hides what's underneath.
Built on Claude's vision capabilities and a lightweight food database for standardized portion estimates. Users can always refine the AI's guess, and the corrected data feeds into both their meal log and the Coach's next conversation.
ML Weight Predictions: Forecast Your Progress
One of FastAI's most powerful features is the weight prediction graph. We trained a lightweight machine learning model on your personal data (fasting duration, meal timing, calories, exercise) to forecast your weight trajectory over the next 4–12 weeks.
class WeightPredictionModel:
def __init__(self):
# Lightweight linear regression + seasonal decomposition
self.historical_weights = [] # daily weigh-ins
self.fasting_hours = [] # daily fasting duration
self.calories_burned = [] # daily exercise/TDEE data
self.calories_logged = [] # daily meal intake
def train(self):
"""Fit model to user's personal data (30+ data points)"""
X = [fasting_hrs, cal_burned, cal_logged, day_of_week]
y = historical_weights
# Use ridge regression to avoid overfitting on small datasets
self.model.fit(X, y)
def predict(self, future_days=28):
"""Generate 4-week weight forecast"""
predictions = []
for day in range(future_days):
# Project forward: fasting hours, typical calorie burn, logged meals
X_future = self._project_features(day)
weight = self.model.predict(X_future)
predictions.append(weight)
return predictions
The model is trained on 30+ data points from the user's own history, not population-wide averages. This makes predictions remarkably accurate and personally motivating: users see "at your current pace, you'll hit your goal by March 15th" with a confidence interval. If the user isn't on track, they can ask Claude, "Why am I not losing weight as fast as I expected?" and Claude can analyze their recent meals, fasting windows, and exercise to identify friction points.
Adaptive Fasting Schedules
FastAI doesn't force a rigid 16:8 schedule on everyone. The app uses machine learning to suggest personalized fasting windows based on your patterns. If you consistently break your fast at 2 PM and eat until 8 PM, the app recommends a 2 PM – 8 PM eating window and 8 PM – 2 PM fasting window (16:8), rather than fighting against your natural rhythm.
Tech Stack: Building a Production AI Health App
Before the deep dive, here's the full production architecture in one diagram — every layer from the App Store down to PostHog event ingest. Open the SVG in a new tab for full detail.
Mobile Architecture
- React Native + Expo: Cross-platform iOS/Android with over-the-air JS updates. Expo Router handles navigation, deep linking, and the typed route registry powering the v1.9 single-render-path onboarding.
- Claude (Sonnet + Haiku) via Convex Serverless: All AI requests go through Convex backend functions — API keys never touch the client, and a tier-router falls back from Sonnet to Haiku for non-critical calls to control cost.
- Voyage 3.5-lite for Memory Embeddings: Powers the Coach's long-term memory and retrieval-augmented context. Cheaper and lighter than larger embedding models — perfect for personal-health memory at scale.
- Convex Database + Crons: Real-time syncing of fasting logs, weight history, meal logs, and AI conversation history, plus nightly memory extraction and RevenueCat reconciliation crons.
- Clerk for Auth (with HIBP): Email/password and Google SSO with HaveIBeenPwned password checking — passwords from known breaches are blocked at sign-up.
- RevenueCat + Superwall + Play Billing: Production paywall live on Android via Play Billing, with Superwall A/B variant assignment and RevenueCat as the receipt-validation source of truth. Server-side reconciliation cron keeps Convex entitlements in sync.
- Firebase Cloud Messaging (FCM): Push notifications for milestone celebrations, gentle re-engagement, and Daily Insight delivery.
- Sentry + PostHog (EU): Errors, releases, and product analytics — privacy-respecting, EU-hosted, with strict noise filtering so user mistakes (wrong password, account-not-found) don't pollute crash dashboards.
Trust, Safety & Engineering Discipline
- Prompt-injection guard on every AI surface: Five Convex AI endpoints share a single hardening helper that wraps user content in unique delimiters and appends an explicit injection guard. Even if a user pastes adversarial text, it can't override the system prompt. Read the audit close-out story →
- Rate limiting + reconciliation crons: Per-user rate limits stop runaway AI cost; nightly RC reconciliation catches any entitlement drift between the store and our DB.
- Account deletion that actually deletes: Cascade across nine tables, validated end-to-end by real testers — not a "request" form that takes 30 days.
Web Presence
- Marketing site at exafabs.ai: Static HTML hosted on GitHub Pages. Fast, dependency-free, no JavaScript framework tax.
- PostHog for product analytics: Funnel, retention, and feature-flag rollouts — EU region, with PII scrubbed before ingest.
Ads Engine API
- FastAPI + Claude AI: Production-grade REST API at api.exafabs.ai with 15 AI-powered endpoints for complete ad campaign automation — audience targeting, keyword strategy, ad copy generation, creative specs, video scripts, funnel mapping, landing page optimization, budget allocation, competitor analysis, A/B testing frameworks, campaign audits, and more.
- Deployed on Railway: Auto-deploys from GitHub with API key authentication, CORS support, and production-grade error handling.
How the AI Coach Works: From Context to Insight
The AI Coach is the secret sauce. Here's the exact flow:
- Data Aggregation (On-Device): When you open the Coach tab, the app assembles your context: last 30 days of fasting logs (with duration and metabolic zones reached), weight history with trend line, meal logs with macronutrients, and your stated goals.
- Context Compression: This data is summarized into a structured prompt (not raw JSON dumps—we use natural language summaries to save tokens and improve Claude's reasoning). Example: "User has been fasting 16:8 for 3 weeks, average weight loss of 1.5 lbs/week. Recent meals show high carb intake on weekends. Goal: lose 15 lbs in 12 weeks."
- Claude API Call (via Convex): The Convex function sends the compressed context + user message to Claude's API. Claude responds with personalized advice, grounded in the user's actual data.
- Streaming & Caching: Response is streamed to the client for instant feedback. Convex caches similar queries (e.g., "why is my weight plateauing?") for 24 hours to reduce API costs.
- Conversation History: All chat turns are stored in Convex so users can revisit past advice, and Claude has full context in multi-turn conversations.
The result: users feel like they have a personal nutrition coach who understands their history, constraints, and psychology—because they do.
What's Next: The Fasting Platform Roadmap
FastAI's MVP is live, but the platform has enormous headroom for innovation:
- Wearable Integration: Connect to Apple Watch, Oura Ring, and Fitbit to track heart rate variability, sleep quality, and activity during fasting windows. Claude uses this data to refine recommendations (e.g., "Your HRV spiked when you did a long fast + hard workout. Let's adjust.").
- Community Challenges: Leaderboards, group fasting challenges, and peer support. Claude can coach users through a 30-day challenge and celebrate milestones.
- Family Plans: Share recipes, fasting schedules, and results with family members. Discounted tier for households.
- Advanced Body Composition Tracking: Integrate DEXA, bioelectrical impedance, or 3D body scanning APIs to track fat loss vs. muscle retention—more nuanced than weight alone.
- Meal Plan Generation: Claude generates weekly meal plans optimized for your fasting window, macronutrient goals, and food preferences. Integration with grocery delivery APIs (Instacart, Amazon Fresh) for one-click ordering.
Ready to Transform Your Fasting Journey?
FastAI Health Coach is LIVE on the iOS App Store (v2.12.2 in review). Android in Closed Testing on Google Play, production access applied. AI-powered fasting coaching, a Coach that remembers your habits, multi-view meal photos, and weight predictions — all in your pocket.
🍎 Download on iOS App Store →