- Current research code (
research-current) — the advanced FitAI research MVP with latent-state and BAI assessment, architecture comparison, PostgreSQL persistence, evaluation workflow, and 54 passing tests. - Stable live demo (
main) — the verified baseline currently deployed on Render:
https://fitai-research.onrender.com/ - Deployment status — the advanced research branch is kept separate from the public demo until its Docker/PostgreSQL deployment and prospective athlete validation are completed.
The live demo demonstrates the stable FitAI baseline.
The defaultresearch-currentbranch contains the latest research architecture referenced in the project presentation.
FitAI is an athlete-specific physiological and latent bioenergetic research system. BAI is an experimental adaptation proxy, not a mitochondrial measurement, diagnosis, or clinically validated score.
Flow: athlete measurements -> canonical 23 features -> trained 23x8 model plus six-encoder research VAE -> BAI/states -> personal/reference comparison -> HJB-inspired safety control -> persisted recommendation.
The trained NumPy artifact is present; a saved hierarchical VAE checkpoint is not. Results disclose this limitation and reduce confidence.
Copy-Item .env.example .env
docker compose up -d db
$env:DATABASE_URL='postgresql://fitai:replace-with-a-strong-password@localhost:5432/fitai'
python manage.py migrateStandard POSTGRES_DB, POSTGRES_USER, POSTGRES_PASSWORD, POSTGRES_HOST,
and POSTGRES_PORT variables are also supported. SQLite is the debug/test fallback.
python -m ml.demo_athlete_assessment --output demo_assessment_output.json
python -m ml.demo_athlete_assessment --input athlete.json --output result.json
python -m compileall -q fitai fitness ml tests
python manage.py makemigrations --check --dry-run
python manage.py check
python -m pytest -qUse --no-persist for a read-only assessment. See
docs/final_mvp_implementation.md, docs/datasets.md,
docs/bai_interpretation.md, and docs/presentation_demo.md.
https://fitai-research.onrender.com/
Press "Get Started" to explore the onboarding flow and interactive physiological AI dashboard. FitAI is an experimental AI-driven physiological fitness platform built with Django, NumPy, OpenCV, and custom mathematical modeling.
The project combines:
- 23 physiological input parameters
- OpenCV-based body photo analysis
- custom NumPy neural networks
- stochastic physiological simulation
- explainability analysis
- dynamic training optimization
- physiological risk modeling
- longitudinal monitoring systems
- temporal prediction tracking
Unlike traditional static fitness calculators, FitAI is designed as a physiologically reactive neural system rather than a fixed rule-based predictor.
The monitoring system tracks longitudinal neural predictions, physiological stability, and adaptive output dynamics over time.
The system dynamically adapts predictions based on:
- HRV
- sleep quality
- emotional stress
- inflammation markers
- alcohol exposure
- cardiovascular state
- hormonal signals
- recovery dynamics
FitAI predicts 8 physiological and fitness-related outputs:
- Calories Burned
- 1 km Run Time
- Cooper Test Distance
- Max Pull-Ups
- Burpees Capacity
- 10 km Run Time
- Waist Circumference Change
- Testosterone Projection
The prediction engine uses:
- custom NumPy neural networks
- hyperbolic tangent activations (
tanh) - softplus physiological output constraints
- momentum optimization
- gradient clipping
- early stopping
- normalization scaling
- physics-based target generation
The neural network is implemented entirely without high-level deep learning frameworks.
The system combines:
- 23 physiological parameters
- OpenCV body proportion analysis from uploaded user photos
Physiological inputs include:
- HRV
- Sleep
- Emotional Stress
- CRP
- Testosterone
- Cortisol
- Resting Heart Rate
- Blood Pressure
- Waist Circumference
- Alcohol Exposure
- Recovery Metrics
- Performance Metrics
The OpenCV analysis module extracts:
- shoulder width
- waist width
- hip width
- body ratios
- physique classification
- training recommendations
A major design goal of FitAI was to avoid creating a random black-box predictor.
The neural system was explicitly designed to demonstrate:
- structured physiological behavior
- parameter sensitivity
- explainability
- stable physiological responses
The model behavior is validated through:
Each physiological feature is independently perturbed by +30% to measure how predictions react.
Local physiological sensitivity is estimated numerically through gradient-based feature analysis.
This means the model does not simply output arbitrary values.
Predictions structurally respond to:
- HRV
- sleep quality
- emotional stress
- CRP inflammation markers
- alcohol exposure
- testosterone
- cardiovascular parameters
Observed physiological behaviors include:
- increased sleep improving recovery-related outputs
- emotional stress worsening endurance predictions
- elevated CRP reducing aerobic performance
- HRV strongly affecting endurance and recovery metrics
- alcohol negatively affecting recovery and hormonal projections
FitAI therefore behaves as a:
rather than a purely statistical estimator.
FitAI includes a built-in explainability pipeline.
The project generates:
feature_importance.json- per-output sensitivity analysis
- relative perturbation analysis
- gradient sensitivity analysis
This allows visualization of:
- which physiological parameters affect each output
- how strongly the neural system reacts
- whether predictions remain physiologically consistent
Examples:
- HRV strongly affects Cooper performance
- Sleep strongly affects Testosterone and recovery
- Emotional stress strongly affects endurance metrics
- CRP affects inflammation-sensitive outputs
FitAI includes an experimental physiological monitoring and evaluation framework designed to track neural network behavior over time.
The platform does not simply display predictions.
Instead, it continuously monitors:
- prediction consistency
- physiological trend dynamics
- temporal stability
- recovery behavior
- cardiovascular response
- hormonal tendencies
- prediction drift across sessions
The evaluation architecture combines two independent systems.
The Stability system measures real neural network behavior across historical onboarding sessions.
Tracked statistics include:
- Stability Mean
- Stability Standard Deviation
- Stability Drift
This layer evaluates:
- prediction consistency
- longitudinal behavior
- temporal variance
- real historical neural response dynamics
Stability metrics are calculated directly from historical prediction records generated by the neural network itself.
The Synthetic system is a separate rule-based physiological simulation layer.
It generates lightweight physiological reference estimates using:
- HRV
- sleep
- emotional stress
- CRP
- weight
- cardiovascular signals
Synthetic metrics include:
- Synthetic Mean
- Synthetic Standard Deviation
- Synthetic Drift
This system acts as:
- a physiological comparison baseline
- a simulation-oriented monitoring layer
- an auxiliary consistency reference
The synthetic layer is implemented independently from the neural network through the synthetic_ground_truth() simulation pipeline.
FitAI therefore implements a:
combining:
- real neural prediction stability
- synthetic physiological simulation comparison
This architecture allows the platform to monitor:
- model drift
- physiological consistency
- behavioral stability
- abnormal prediction deviations
- recovery trend evolution
FitAI computes an experimental:
derived from:
- stability drift
- synthetic drift
- historical prediction variance
The score acts as a lightweight behavioral monitoring signal for neural prediction consistency.
The purpose of the system is not medical validation, but:
- neural behavior monitoring
- physiological trend visualization
- experimental AI evaluation
- recovery-oriented tracking
FitAI includes dynamic physiological dashboards built with:
- Django
- JavaScript
- Chart.js
The visualization layer provides:
- prediction timelines
- historical physiological graphs
- model stability charts
- explainability displays
- monitoring dashboards
- temporal drift visualization
-
FitAI contains a stochastic emotional forecasting system inspired by Itô stochastic differential equations.
Mathematical form:
The system models:
- emotional stress drift
- nonlinear alcohol influence
- volatility growth
- stochastic instability
- mean reversion dynamics
Alcohol exposure increases:
- stress drift
- emotional instability
- volatility of future emotional states
This module introduces stochastic physiological modeling into the platform.
FitAI includes a physiological weekly training optimizer.
The optimizer dynamically adjusts:
- calorie deficit
- weekly HIIT count
- recovery load
- cardiovascular stress
- recovery penalties
The optimizer considers:
- HRV
- sleep quality
- alcohol exposure
- blood pressure
- age
- training load
- recovery penalties
FitAI includes a simplified physiological control model inspired by Hamilton–Jacobi–Bellman optimization ideas.
The system estimates:
- training risk
- optimal training intensity
- short-term physiological trajectory
The model simulates:
- HRV dynamics
- sleep recovery dynamics
- blood pressure response
- physiological cost minimization
Simplified objective form:
where:
- (X_t) → physiological state
- (u_t) → training intensity/control signal
The system searches for lower-risk physiological trajectories across time.
FitAI uses a dual-dataset architecture.
edited_23_params_realistic.csv
Contains:
- engineered physiological variables
- observable biomedical signals
- synthetic physiological feature generation
- stable production-oriented training data
Generated using:
- pandas
- NumPy
- physiological feature engineering
edited_23_params_realistic_latent.csv
Experimental hidden-state physiological dataset containing:
- latent_energy
- latent_stress
- latent_recovery
- latent_rhythm
The latent pipeline introduces:
- hidden-state physiological representations
- normalized latent synthesis
- experimental recovery modeling
- future research-oriented architecture
FitAI therefore supports architectural switching between:
- observable physiological learning
- latent hidden-state augmented learning
The latent dataset pipeline is reserved for future experimental research and advanced model extensions.
Main architecture:
- Input: 23 physiological parameters
- Embedding layer: 48 neurons
- Hidden layer: 32 neurons
- Output layer: 8 physiological outputs
Activation functions:
- Hidden layers → tanh
- Output constraints → softplus + tanh
Additional features:
- momentum optimization
- gradient clipping
- early stopping
- normalization scaling
- physiological output constraints
The trained model is stored as:
trained_fitness_model_simple.pkl
Serialized components include:
- neural network weights
- embeddings
- normalization statistics
- output scaling state
- validation metrics
The utility script:
check_model.py
The project includes:
- physiological prediction tests
- behavioral consistency tests
- deterministic inference tests
- validation tests
- Django form validation tests
Implemented using:
- pytest
- Python
- Django
- NumPy
- pandas
- OpenCV
- SQLite
- custom NumPy neural networks
- explainability analysis
- stochastic physiological simulation
- sensitivity analysis
- physiological monitoring systems
- Docker
- docker-compose
- pytest
- HTML
- CSS
- JavaScript
- Chart.js
FitAI/
│
├── data/
│ ├── edited_23_params_realistic.csv
│ │ # Expanded physiological dataset generated with pandas feature engineering
│ │
│ ├── edited_23_params_realistic_latent.csv
│ │ # Physiological training dataset augmented with hidden-state latent features
│ │
│ └── gym_members_exercise_tracking.csv
│ # Original dataset downloaded from Kaggle
│
├── fitai/
│ ├── settings.py
│ └── urls.py
│
├── fitness/
│ ├── migrations/
│ │
│ ├── static/
│ │ ├── css/
│ │ └── js/
│ │
│ ├── templates/
│ │ ├── base.html
│ │ ├── history.html
│ │ ├── metrics.html
│ │ ├── onboarding.html
│ │ ├── results.html
│ │ └── update.html
│ │
│ ├── apps.py
│ ├── forms.py
│ ├── urls.py
│ │
│ └── views.py
│ # Core Django application logic integrating onboarding,
│ # physiological prediction, explainability, training optimization,
│ # risk analysis, photo analysis, and model stability metrics
│
├── media/
│ └── photos/
│
├── ml/
│ ├── models/
│ │ ├── check_model.py
│ │ │ # Utility script for converting serialized .pkl model contents
│ │ │ # into a human-readable parameter overview
│ │ │
│ │ └── trained_fitness_model_simple.pkl
│ │ # Serialized trained physiological prediction model
│ │ # with weights and normalization state
│ │
│ ├── emotional_drift.py
│ │ # Stochastic emotional stress forecasting using an
│ │ # Itô-process-inspired physiological drift model
│ │
│ ├── feature_importance.json
│ │ # Feature importance and explainability analysis for model outputs
│ │
│ ├── fit_model_core.py
│ │ # Custom NumPy neural network with physiological output constraints,
│ │ # momentum optimization, early stopping, and feature importance analysis
│ │
│ ├── photo_analysis.py
│ │ # OpenCV-based body proportion analysis and physique classification
│ │ # from user photos
│ │
│ ├── preprocess_dataset.py
│ │ # Physiological dataset expansion and synthetic feature engineering
│ │ # pipeline using pandas and NumPy
│ │
│ ├── preprocess_latent.py
│ │ # Latent physiological feature generation and hidden-state
│ │ # normalization pipeline
│ │
│ ├── train_model.py
│ │ # NumPy neural network training pipeline with physics-based
│ │ # target generation, validation, serialization,
│ │ # and explainability analysis
│ │
│ ├── training_history.json
│ │ # Training configuration, validation metrics, and model metadata
│ │
│ ├── training_optimizer.py
│ │ # Physiological weekly training load optimizer with recovery
│ │ # and cardiovascular risk adjustment
│ │
│ └── training_risk.py
│ # Simplified Hamilton–Jacobi–Bellman-inspired physiological
│ # training risk and intensity optimization model
│
├── tests/
│ ├── test_model.py
│ │ # Pytest-based validation suite for physiological prediction
│ │ # consistency and behavioral sanity checks
│ │
│ └── test_views.py
│ # Input validation tests for Django form and physiological data handling
│
├── .dockerignore
├── .gitignore
├── db.sqlite3
├── docker-compose.yml
├── Dockerfile
├── main.py
├── manage.py
├── pytest.ini
└── requirements.txt
Experimental visualization of mitochondrial energy behavior, ATP efficiency, recovery energetics, and physiological adaptation mechanisms.
FitAI explores future physiological modeling concepts involving:
- ATP production efficiency
- recovery energetics
- mitochondrial adaptation
- oxygen utilization
- latent energy dynamics
- physiological resilience modeling
This direction is part of the platform’s long-term research architecture focused on hidden physiological state modeling and adaptive biological intelligence systems.
Future directions include:
- latent-state physiological learning
- hidden recovery state modeling
- nonlinear regression extensions
- sinusoidal physiological dynamics
- quaternion-based neural layers
- physiological hidden-state embeddings
- advanced optimal-control systems
- online adaptive learning
- stochastic recovery simulation
FitAI was designed not as a generic calorie predictor, but as an experimental physiologically reactive AI system.
The project attempts to combine:
- machine learning
- physiological modeling
- explainability
- stochastic mathematics
- optimal control ideas
- recovery dynamics
- hidden-state modeling
- longitudinal monitoring
- physiological simulation
into a single extensible research-oriented architecture.