TypeScriptAIMobile

ZeroCoach

Since 2026 - present

An open-source, privacy-first AI health coach with offline mobile LLM inference, visual nutrition tracking, wearable sync, and adaptive lifestyle.

Visit Live Site →GitHub Repository

ZeroCoach

Website GitHub Release Download APK ![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)

ZeroCoach is an open-source, privacy-first mobile fitness and athletic AI coaching application powered entirely by on-device and local AI models. Built with React Native and Expo, it delivers science-backed workout programming, adaptive nutrition targets, and real-time coaching—with zero cloud latency and absolute data privacy.

Visit the official website at zerocoach.tech.

🎬 Demo Video

ZeroCoach Video Demo

▶️ Watch the Full HD Demo on YouTube: https://youtu.be/Ns-7ikl16K0

Featuring on-device Gemma 2 GGUF inference streaming at ~38 tok/s, multimodal food plate vision decomposition & micronutrient analysis, RepDB adaptive workout engine, and unified 7-day athletic telemetry.

📸 App Preview

💬 Coach Chat 🥗 Nutrition Tracker 🏋️ Adaptive Workout
Coach Chat Nutrition Tracker Adaptive Workout
Gemma 2 On-Device Coach
Real-time athletic guidance & dual persona		 Smart Nutrition & Macros
Gemini Vision, NutriSight & 11 RDAs		 Adaptive Workout Hub
RepDB 643 exercises & Date Navigation

📈 Performance & Trends ⚙️ On-Device Engine
Performance & Trends On-Device Engine Settings
Multi-Domain Analytics
0–100 composite score & Health Connect		 Hardware Acceleration
Quantized GGUF execution & local endpoints

✨ Key Features

  • ⚡ AI Chat (Agent Mode with Action Harness): Autonomous tool execution harness that reads and writes across all 8 core app domains (telemetry, food history, food DB, workout history, workout plan, RepDB, athlete settings, and athlete memory) in real-time.
  • 🧠 On-Device Athlete Memory Layer: Local-first long-term memory engine with ADD / UPDATE / DELETE / NOOP reconciliation, dense vector similarity, and bilingual fact extraction—remembers favorite foods, dietary restrictions, favorite workout skills, injuries, and personality preferences across sessions.
  • 📱 100% On-Device AI Inference: Runs quantized LLM models (e.g. Google Gemma 2) locally with hardware acceleration—zero cloud latency, zero API costs, and absolute data privacy.
  • 🔌 OpenAI-Compatible Local Endpoints: Connect seamlessly to local network LLM runners like Ollama, LM Studio, vLLM, or desktop llama.cpp servers.
  • 🔄 Symmetric UI & AI Access Parity: 100% parity between direct App UI manipulation and conversational AI commands—log meals, swap workout exercises, modify macro targets, and view wearable biometrics conversationally or through screens.
  • 💬 Conversational Athletic Coach: Real-time direct & professional sports-science coaching, auto-regulating advice based on live recovery metrics, persistent training plans, and nutrition.
  • 🏋️ Goal-Driven Adaptive Training & Persistent Plans: Automatically generates personalized workout splits based on readiness scores, equipment (Gym vs. Bodyweight), and athlete skill level, stored persistently by date.
  • 📅 Calendar Date Navigation: Seamless date navigation across both Workout and Nutrition modules—review past sessions, modify upcoming routines, or plan ahead with instant 1-tap jump to today.
  • 📚 RepDB Exercise Library: Integrated catalog of 643 science-backed exercises with start/peak pose previews, customizable movement swapping, and skipping.
  • 🥗 Smart Nutrition & Multimodal Vision: Dual-mode camera scanner featuring Gemini 3.5 Flash Lite food plate analysis and Open Food Facts NutriSight label extraction alongside 11 micronutrient RDA bio-scores.
  • 📈 Unified Athletic Analytics: Multi-domain bio-history tracking recovery, workout volume, and nutrition with composite scoring and SVG trendline charts.
  • ⌚ Universal Wearable & Health Sync: Direct on-device sync with Android Health Connect and Apple HealthKit to aggregate sleep, HRV, heart rate, and steps across Garmin, Apple Watch, Whoop, Oura, Samsung, and more.
  • 🎨 Privacy-First Obsidian Interface: Dark-mode glassmorphic aesthetic optimized for fluid mobile interactions and system safe insets.

🏗️ Architecture & Tech Stack

ZeroCoach is engineered for privacy, speed, and resilient offline execution with complete symmetric access parity:

📘 Full Architecture Specification: See **docs/ARCHITECTURE.md** for detailed Mermaid diagrams, sequence flows, tool schemas, and token-saving patterns.

Layer Technology
Framework React Native with Expo (TypeScript)
Agent Action Harness Autonomous multi-turn Tool Execution Dispatcher with live action status pills and hybrid tool calling protocol
On-Device Inference llama.rn (React Native bindings for llama.cpp) running quantized GGUF models
Local LLM Providers OpenAI-compatible SSE streaming client supporting Ollama, LM Studio, and vLLM
Web Inference Engine Automatic deterministic simulation engine when running in browser environments
Exercise Database Open Exercise Repository adapted from RepDB (643 movements with verified CDN pose assets)
Nutrition Databases Open Food Facts API + USDA FoodData Central whole foods dataset + NutriSight token extraction
State Management Zustand with AsyncStorage local persistence (historyByDate, plansByDate, logsByDate)
Styling & Design Sleek Cyber-Athletic Dark Mode (#080B10), HSL accents, and glassmorphic micro-components
Icons Lucide React Native
Marketing Website Astro + Tailwind CSS in /web (deployed on Render at zerocoach.tech)

📊 8-Domain Symmetric Read/Write Parity

Domain App UI Access AI Agent Harness Access
1. Fitness Telemetry View Readiness (0–100), HRV, resting HR, sleep, steps, and sync wearables Read: get_telemetry_metrics()
Action: sync_fitness_telemetry()
2. Food History Browse meal logs by date, log meals, delete items, update macro targets Read: get_food_history()
Write: log_food_item(), delete_logged_food(), update_daily_macro_targets()
3. Food Database Search whole foods (USDA), query packaged goods (Open Food Facts), scan barcodes Read: search_food_database(), get_food_details()
4. Workout History Review past workout sessions, volume, and sets by calendar date Read: get_workout_history()
Write: log_completed_workout()
5. Workout Plan View adaptive daily splits, swap/skip exercises, schedule future routines Read: get_workout_plan()
Write: update_workout_plan(), swap_plan_exercise()
6. Workout Database (RepDB) Browse 643 movements, filter by muscle/equipment, view poses & cues Read: search_exercises(), get_exercise_details()
7. ZeroCoach Settings Update athlete profile, goals, weight/height, switch coach persona, language Read: get_user_settings()
Write: update_user_settings()
8. Athlete Memory Browse, filter, add, or delete categorized personal memories in Settings Read: get_memories()
Write: save_memory(), remove_memory()

📱 App Navigation & Modules

1. 💬 Coach (CoachChatScreen)

  • Always-On Action Harness: Automatically calls tools to read/write food logs, workout plans, and biometrics directly from natural chat.
  • Live Activity Feed: Displays real-time tool execution pills (e.g. ⚡ HARNESS: LOGGING 200G CHICKEN BREAST...) and verification badges upon completion.
  • Formatted response rendering for structured lists, kinetic cues, bold titles, and biomechanical guidance.
  • Synchronized with active wearable telemetry (HRV, resting heart rate, sleep duration, calorie burn) and today's logged nutrition totals & remaining macros.
  • Context-aware coaching chips:
    • Suggest Food, Suggest Workout Plan, Adjust Macros, Readiness Review, Bodyweight Cues.
    • Deep-link action pills ("Log Food in Nutrition Tracker", "View Workout & Progressions").

2. 🥗 Nutrition (NutritionScreen)

  • Macro Progress Header: Real-time calorie gauge, remaining energy budget, and progress bars for Protein, Carbs, Fat, and Dietary Fiber.
  • Date Navigation: Easily browse, edit, and log meals across any historical date or jump to today.
  • Micronutrient Bio-Profile Scorecard: Expandable RDA tracker for Vitamins (A, C, D, B12) and Minerals (Iron, Calcium, Magnesium, Potassium, Sodium limit, Zinc).
  • Meal Sections: Structured daily tracking across Breakfast, Lunch, Dinner, and Snacks.
  • Log Food Modal:
    • Database Search: Search whole foods (USDA) and packaged products (Open Food Facts).
    • Barcode Scanner: Scan barcodes or enter codes with instant OFF product lookup.
    • Snap / NutriSight Label Scanner: Dual-mode camera scanner featuring Open Food Facts NutriSight (openfoodfacts/nutrition-extractor LayoutLMv3 token schema) to extract exact nutrition tables (calories, protein, carbs, fat, fiber, sodium, and serving grams) from packaging labels, alongside multimodal cooked meal plate decomposition via Gemini 3.5 Flash Lite.
    • Portion Review Sheet: Adjust grams/servings with real-time nutrient recalculation before logging.

3. 🏋️ Workout (ProgressionsScreen)

  • Date Navigation: Full parity with nutrition tracking—navigate between previous days' recorded workouts and future planned splits.
  • Training Goal & Skill Level Selectors: Tailor routines for Strength, Calisthenics, Fat Loss, Cardio, or Longevity across Beginner, Intermediate, or Advanced tiers.
  • Workout Split Hub: Select Full Body, Push, Pull, Legs, Upper, Lower, Core & Mobility, Calisthenics Skills, or Cardio Engine.
  • Coach Suggestion Banner: 1-tap apply for CNS readiness-tailored split recommendations.
  • Adaptive Daily Exercise Plan:
    • Complete routine with dynamic warmups, primary compound lifts, secondary volume, and cooldowns.
    • RepDB movement specifications: target sets/reps, rest intervals, equipment, target muscles, and MET burn.
    • Skip / Restore: Skip exercises dynamically without distorting completion progress.
    • Swap Movement: Instant replacement with any exercise targeting the same muscle group.
    • Add Skill / Exercise from RepDB: Search and select from 643 exercises (with quick filters for Pull-ups, Dips, Squats, etc.).
    • Kinetic Pose Previews: Exercise detail modal showing start and peak pose images.

4. 📈 Analyze & Telemetry (AnalyzeScreen)

  • Unified Athletic Performance Score: Daily 0–100 composite score combining Central Nervous System (CNS) recovery (35%), workout adherence (35%), and nutrition target accuracy (30%).
  • Interactive Multi-Domain Bio-History: Filter across 7-day, 14-day, and 30-day time ranges with smooth SVG trendline charts and volume bar charts.
  • Dynamic CNS Readiness Gauge: Evaluates recovery status based on HRV, resting heart rate, sleep duration, and deep sleep architecture.
  • Native OS Health Hub Bridge: Connects locally to Android Health Connect or Apple HealthKit to read biometrics from paired wearables (Garmin, Apple Watch, Samsung, Whoop, Oura, Pixel, Xiaomi, Strava) with zero remote API latency.
  • Domain Breakdown: Dedicated trend tabs for Workout (sessions, active burn, set volume), Nutrition (calorie compliance, macro balance), and Recovery (sleep quality, HRV).
  • Smart Scale Auto-Sync: Automatically updates athlete body weight and recalibrates daily nutrition targets whenever a new weigh-in is detected.

5. ⚙️ Settings & Engine (ModelManagerScreen)

  • Inference Provider Configuration:
    • On-Device: Direct quantized GGUF execution via llama.rn on iOS/Android, with automatic browser fallback for web previews.
    • OpenAI-Compatible: Connect to local desktop or LAN endpoints (e.g., http://192.168.1.100:11434/v1 for Ollama).
  • Coach Memory (On-Device Vector Layer): View, filter, add, and delete categorized personal memories (Food & Diet, Workout & Skills, Injuries & Limits, Personality & Style).
  • Local Storage Management: One-tap clearance for chat transcripts, coach memories, or full athlete biometric data wipe.

🚀 Getting Started

Prerequisites

  • Node.js: >= 18.0.0
  • npm or yarn
  • Mobile Development: Xcode (iOS) or Android Studio / Android SDK (Android)
  • Web Development: Modern Chromium, Safari, or Firefox browser

Installation

# 1. Clone the repository
git clone https://github.com/ntsd/zerocoach.git
cd zerocoach

# 2. Install dependencies
npm install

Running Locally

# Start the Expo development server
npm start

# Run in Web browser (Instant preview with automated web engine)
npm run web

# Run on Android (Physical device or emulator)
npm run android

# Run on iOS (Physical device or simulator - macOS only)
npm run ios

🌐 Marketing Website (/web)

The official marketing website for ZeroCoach is located in /web and hosted on Render at zerocoach.tech.

Local Development

cd web
npm install
npm run dev

Production Build

cd web
npm run build
npm run preview

Deployment on Render

ZeroCoach includes a root `render.yaml` Blueprint configured for zero-config Render static site deployments:

  • Build Command: npm install && npm run build
  • Publish Directory: dist
  • Root Directory: web
  • Custom Domains: zerocoach.tech and www.zerocoach.tech

📦 Standalone Android APK (No Play Store Required)

Pre-built signed release APKs are available on the GitHub Releases page for immediate download.

To build the APK locally from source:

# Build standalone release APK (bundles JS, assets, and native llama.rn libraries)
npm run build:apk

# Generated APK output:
# android/app/build/outputs/apk/release/app-release.apk

# Sideload directly to a connected Android device:
adb install -r android/app/build/outputs/apk/release/app-release.apk

⚡ Over-The-Air (OTA) Updates (iOS & Android)

ZeroCoach supports Over-The-Air (OTA) hotfixes via expo-updates. Bug fixes, coach prompt refinements, UI improvements, and calculation updates can be deployed directly to existing Android and iOS installs without rebuilding binaries:

# Push an instant OTA hotfix to both Android and iOS installs
npm run update:publish -- --message "Fix macro calculation and prompt tuning"

Users can also tap "CHECK FOR OTA UPDATES" anytime in Settings → Over-The-Air (OTA) Hotfix Channel to check for and apply updates on-demand.

🧠 Model Configuration & GGUF Support

When using on-device inference on physical devices:

  1. Primary Recommended Baseline: Google Gemma 2 2B Instruct (gemma-2-2b-it-Q4_K_M.gguf, ~1.64 GB). The official default on-device engine for robust multi-turn coaching, biomechanical cues, and structured nutrition planning.
  2. Experimental Compact Model: ZeroCoach Athletic Intelligence (zerocoach-athletic-q4_0.gguf, 91.7 MB). An experimental 135M parameter fine-tune quantized to Q4_0, built for ultra-lightweight offline mobile tests within GitHub's 100 MB push limit.
  3. Desktop / LAN Serving: For zero mobile RAM consumption, run ollama run gemma2:2b or LM Studio on your local network and point Settings → OpenAI-Compatible Provider to your host IP.

🤖 Model Fine-Tuning & Vector Embeddings Pipeline

ZeroCoach includes a complete, autonomous machine learning pipeline located in `/ml` to fine-tune on-device language models (e.g. Google Gemma 2 2B/9B, SmolLM, Qwen) and pre-compute dense vector embeddings for instantaneous sub-millisecond retrieval.

1. Dataset Generation (/ml/scripts/generate_dataset.py)

Synthesizes 1,075 verified athletic coaching pairs across 8 core domains formatted in native Gemma 2 chat tokens (<start_of_turn>user...<end_of_turn><start_of_turn>model...<end_of_turn>) and Alpaca JSON:

  • Biomechanical Compound Lifts & Kinetic Cues (Squat, Bench, Deadlift, Row, OHP)
  • Sustainable Joint-Friendly Adaptations & Kinetic Recovery
  • Calisthenics Leverage Milestones (RepDB bodyweight standards)
  • Autoregulation for Low Readiness (<60 / Deload) & High Readiness (>=80 / Overload PRs)
  • Precision Macronutrient Optimization (Leucine threshold >= 2.7g)
  • Movement Swapping & Pain Mitigation
# Generate the dataset (outputs to ml/data/)
./ml/venv/bin/python3 ml/scripts/generate_dataset.py

2. Dense Vector Embeddings (/ml/scripts/generate_embeddings.py)

Computes 384-dimensional normalized dense vectors for all 643 RepDB movements and USDA whole food items using sentence-transformers/all-MiniLM-L6-v2. Enables offline semantic search, movement swapping, and smart meal lookups with zero network calls:

# Generate vector indices (JSON & safetensors)
./ml/venv/bin/python3 ml/scripts/generate_embeddings.py

# Test local semantic vector retrieval
./ml/venv/bin/python3 ml/embeddings/vector_search.py

3. Local LoRA Fine-Tuning (/ml/scripts/finetune.py)

Trains parameter-efficient LoRA adapters using Apple Silicon MPS (Metal Performance Shaders) or CUDA acceleration:

# Train on device with MPS GPU acceleration
./ml/venv/bin/python3 ml/scripts/finetune.py --epochs 2 --batch_size 4 --grad_accum 2

# Checkpoint output: ml/output/zerocoach_lora_adapter/

4. Local Inference Verification (/ml/scripts/test_inference.py)

Runs the fine-tuned LoRA adapter locally across real telemetry scenarios (readiness scoring, fatigue deloads, macro suggestions):

./ml/venv/bin/python3 ml/scripts/test_inference.py

5. Web Distribution on zerocoach.tech

All datasets, adapters, and embedding tensors are staged for direct public download from zerocoach.tech/models/:

  • zerocoach_dataset.jsonl (Full 1,075 sample JSONL)
  • zerocoach_alpaca.json (Alpaca format)
  • zerocoach_embeddings.safetensors (Binary normalized embeddings)
  • adapter/adapter_model.safetensors (Trained LoRA weights)
  • model_manifest.json (SHA-256 verified release manifest)

📄 License

ZeroCoach is released under the Apache License 2.0. See the LICENSE file for details.