Athletic Intelligence.
Running Directly On Your Device.
Zero subscription lock-in. Zero cloud servers. Powered by quantized Gemma 2 running locally on mobile hardware with llama.rn, adaptive RepDB training splits, multimodal food vision, and real-time wearable bio-telemetry.
See ZeroCoach in Action.
Watch our 1080p full production walkthrough featuring real on-device Gemma 2 inference, multimodal meal decomposition, and kinetic exercise tracking.
Interactive App Features & Module Showcase.
Tap between screens below to explore ZeroCoach's obsidian glass interface, biometric feeds, and on-device athletic tools.
Coach Chat
Gemma 2 On-Device Athletic Coach
Private conversational AI delivering biomechanical form cues, nutritional adjustments, and CNS recovery guidance. Runs 100% locally with live token streaming.
Core System Features & On-Device AI Architecture.
Every algorithm, model inference, and biometric calculation happens on your hardware. Zero cloud dependencies, zero data brokerage.
100% On-Device Gemma 2 Coach
ZeroCoach executes quantized Gemma 2 models directly on your smartphone via llama.rn and native llama.cpp. Powered by an on-device Vector Memory Layer, your coach remembers your favorite foods, target calisthenics skills, injuries, and personality across sessions.
Smart Nutrition & NutriSight
Snap a food plate for multimodal decomposition using Gemini 3.8 Flash, or point your camera at nutrition fact labels with Open Food Facts NutriSight token extractors. Includes USDA whole foods and 11 micronutrient RDA bio-gauges.
RepDB Workout Engine
Access 643 exercises from the verified RepDB repository with kinetic start and peak pose imagery. Dynamically adjust volume to your readiness, skip or swap movements on the fly, and review past or future training splits with dedicated calendar date navigation.
Unified Athletic Performance Analytics
Connect directly on-device with Android Health Connect and Apple HealthKit to unify HRV, resting heart rate, sleep architecture, and burn across Garmin, Apple Watch, Whoop, Oura, Samsung, and Pixel wearables. Generates a daily 0–100 composite score with 7D, 14D, and 30D trendlines.
Architectural Comparison: ZeroCoach vs. Cloud Fitness Apps.
Why running AI models on your mobile silicon changes the game for athletic training and personal health privacy.
| Capability | ZEROCOACH (On-Device) | Typical Cloud Fitness Apps |
|---|---|---|
| Model Execution | 100% On-Device (llama.rn & GGUF) | Remote Cloud Servers (OpenAI/AWS) |
| Inference Latency | 0ms Network Latency (Immediate) | 800ms – 2,500ms Cloud Round-trips |
| Offline Usability | 100% Functional Without Internet | Fails Without Live Internet Connection |
| Biometric Data Ownership | Stored Locally On Your Device | Monetized or Stored on Cloud Datacenters |
| Subscription Cost | $0 / Lifetime Free (Apache 2.0) | $15 – $35 / Month Recurring |
| Custom LLM Choice | Any Quantized GGUF or Local LAN Server | Locked to Provider's Proprietary Model |
| Wearable Integration | Direct OS Health Connect & HealthKit | Third-party Cloud API Integrations |
| Athlete Memory Layer | 100% Local Vector Engine (Foods, Skills, Injuries) | Stateless Prompts or Cloud-Profile Tracking |
System Specifications & Open Engineering Architecture.
A resilient, fully open-source architecture designed with 100% symmetric parity between the Mobile UI and AI Agent Action Harness.
React Native + Expo
Built with TypeScript for smooth 60fps performance on iOS and Android. Obsidian theme tokens, glassmorphism, and responsive safe insets.
Intent Router + Tool Harness
Sub-millisecond bilingual intent classifier (EN/TH) scoping tool declarations down to ~75–140 tokens while keeping ambient athlete memory permanently active.
llama.rn & Session-Scoped Vision
Native C++ execution with ARM64 NEON & Apple Metal. 4,096-token KV capacity with on-demand vision projector lifecycle saving 600 MB heap during text coaching.
RepDB Verified Repository
643 science-backed movements with start and peak pose kinematics, dynamic warmup builders, and CNS readiness load auto-regulation.
Gemini Vision & NutriSight
Plate meal decomposition via Gemini 3.8 Flash, Open Food Facts barcode lookup, and NutriSight label extraction for 11 micronutrients.
Health Connect + Zustand
OS health aggregators reading resting HR, HRV, and sleep. All logs stay encrypted in local AsyncStorage with 1-tap instant data wipe.
App UI & AI Agent Parity Matrix
| Domain | Access | App UI Access | AI Agent Harness Access |
|---|---|---|---|
| Fitness Telemetry | Read / Action | View Readiness (0–100), HRV, resting HR, sleep, steps, and sync wearables | Read biometrics via get_telemetry_metrics(), trigger sync with sync_fitness_telemetry() |
| Food History | Read / Write | Browse meal logs by date, log meals, delete items, update macro targets | Read food history, log food items, delete entries, and update macro targets |
| Food Database | Read | Search whole foods (USDA), query packaged goods (Open Food Facts), scan barcodes | Search USDA & Open Food Facts databases and fetch nutrient profiles by barcode |
| Workout History | Read / Write | Review past workout sessions, volume, and sets by calendar date | Read workout history across recent days and log finished workout sessions |
| Workout Plan | Read / Write | View adaptive daily splits, swap/skip exercises, schedule future routines | Read daily workout plans, update full routines, and swap exercises from RepDB |
| Workout DB (RepDB) | Read | Browse 643 movements, filter by muscle/equipment, view poses & cues | Search 643 movements and retrieve kinetic cues, biomechanics, and pose metadata |
| ZeroCoach Settings | Read / Write | Update athlete profile, goals, weight/height, switch coach persona, language | Read athlete profile settings and update goals, body weight, coach tone, or language |
| Athlete Memory | Read / Write | View, filter, add, or delete categorized memories (food, skills, injuries, personality) | Reconcile facts (ADD/UPDATE/DELETE) via save_memory(), get_memories(), remove_memory() |
Open Athletic AI Models & Engineering Specifications.
We believe athletic intelligence belongs on your personal device. Download our fine-tuning datasets, LoRA adapters, and dense vector embeddings for 100% offline coaching.
Gemma 2 & ZeroCoach GGUF
Default Baseline: Google Gemma 2 2B Instruct (Q4_K_M, 1.64 GB) for full multi-turn athletic reasoning.
Experimental: ZeroCoach Athletic 135M (Q4_0, 91.7 MB) for ultra-compact offline mobile demos.
Athletic SFT Dataset
1,075 verified training pairs across 8 core athletic domains: strict kinetic mechanics, calisthenics progressions, deload/overload readiness rules, and leucine thresholds.
RepDB Vector Store
384-dimensional normalized dense vectors for all 643 RepDB exercises & whole foods. Complete LoRA fine-tuning and GGUF quantization scripts in the repo.
Download ZeroCoach v1.1 for Mobile.
Install the pre-compiled signed release APK on your Android phone, or clone the repository to run on iOS or local emulators.
adb install -r ZeroCoach-v1.1.0.apk