KINETICA AI

Clinical AI that starts every query from scratch is brittle. Clinicians need persistent context — patient profile, clinical rules, evidence base — injected into every model call.

Persistent orchestration layer in front of the LLM. Each query is enriched with patient profile, clinical rules, RAG evidence (1,880 chunks in ChromaDB), and ALMA L1 axioms before any model sees it. The model is interchangeable. IO3 is not.

9-node LangGraph graph with human-on-loop interrupt at every gap. Full reasoning audit trail. Designed for EU AI Act compliance — clinician decides at every uncertainty, never the model alone.

Patients with complex chronic conditions can't predict symptom flares. Crashes arrive without warning, 24–72h after the trigger.

N=1 longitudinal study: 207 nights of nocturnal HRV from a consumer wearable. Five independent models, each selecting its own features via forward selection across 13 candidates. Validated on 61 prospective pairs with LOO-CV.

AUC 0.83 (autonomic dysfunction). Headline metric uses nocturnal RMSSD — physiologically coherent, not previously reported in N-of-1 longitudinal Lyme/ME-CFS literature. All code and data public.

LLMs in clinical contexts need guardrails that are not just prompt tricks. Prompt-based safety fails silently.

Two layers. L1: pre-generation injection of 5 axioms (Conciencia, Claridad, Límite, Pragmatismo, Cuidado) into every model call — deterministic, zero runtime overhead. L2: post-generation evaluation, currently being redesigned. Known limitation publicly documented: in current L2, Haiku evaluates its own outputs and RLHF dominates over the axioms.

L1 operative in production. L2 redesign in progress. Decisions emit APPROVE / REWRITE / SILENCE — clinician always decides. Three structural bugs publicly documented at architecture page. Honest about what works and what does not.