End-to-end clinical AI, engineered on real longitudinal physiology. Not abstract benchmarks.
End-to-end · open data · reproducible
Wearable physiology and prospective symptom diaries land in one versioned, longitudinal pipeline. The same archive feeds every downstream model, from predictors to agent reasoning and safety audits. Devices and features may change over time; the contract with the data does not.
Models trained on one patient's deep, longitudinal physiology, not on cohorts. Each one targets a different clinical signal. The same scaffold accepts new targets, new sensors, new validation methods.
Clinical reasoning, guarded by design. An agent loop handles context, tools and multi-step thought, audited end to end. A deterministic safety layer scores every response against clinical boundaries before it reaches a clinician.
Every piece of the system is public. Each card below opens to code and a reproducible run, with claims anchored in peer-reviewed sources.
Pipeline Polar & Symptoms
Polar exports and prospective symptom diaries flow into a clean longitudinal dataset, ready for modelling. Advanced HRV features, deterministic nightly jobs, every level versioned L0→L6.
View pipeline →ANS Predictor
Research prototype. Multi-target models estimate symptom burden from nocturnal HRV and diary-linked physiology. Leave-one-out validation, bootstrap CIs, all built on the same Polar pipeline.
AUC 0.84 autonomic · 0.92 severity · n=61 · N-of-1
View predictor →Sleep Quality Predictor
Research prototype. Sleep quality treated as its own clinical signal, not a byproduct. Same cleaned physiology, focused on how nocturnal structure and autonomic patterns track perceived sleep degradation and recovery.
AUC 0.77 sleep quality · same physiological foundation
Explore sleep model →Cross-Predictor Convergence
Where two independent models agree. ANS and Sleep each selected nocturnal RMSSD as their top fatigue feature, on their own. This page shows feature overlap, AUC on shared days and day-level probability agreement.
r=0.66 · 79% agreement · 1 shared feature · n=42 shared days
Explore convergence →IO3 Clinical Agent
Research prototype. LangGraph agent that orchestrates Anthropic models, clinical rules and retrieval for guarded chronic-care reasoning. One audited loop, human-on-loop control, traceable session logs.
View architecture →ALMA Safety & Evaluation
Research prototype. Deterministic safety layer screening agent responses for pharmacological risk, diagnostic overreach, false urgency and scope violations. Evaluated on a 30-case clinical test set with per-severity metrics and millisecond-level latency.
See safety layer →Clinical Knowledge & RAG
Curated knowledge base of 1,880 audited chunks across HRV, PEM, osteopathy, neurodynamics and portfolio content. RAG pipelines are tested on a 20-question benchmark with 0.85 retrieval accuracy overall.
Explore knowledge stack →Open Research Repository
Public repo hosting the Polar pipeline, predictor code, notebooks and study materials behind Kinetica's current research line. Structured for reproducible runs, not marketing screenshots.
View on GitHub →Kinetica AI is built by Alfonso Navarro. Physics at Universidad de Granada, with postgraduate work in biomechanics. Trained in osteopathy at UAB. Ten years of independent clinical practice in the Pyrenees: complex musculoskeletal and neuromechanical cases, high-performance athletes, mountain-sport injuries. Two years of acute COVID hospital care in Vielha during the pandemic. He is also the patient. The system is engineered from real physiological uncertainty, not benchmark chasing. Wearable monitoring, longitudinal symptom data and interpretable architectures for clinical AI.
Clinical AI consulting · Research-grade HRV analysis · AI model evaluation