Pipeline Overview

How ConceptualHealth.AI Works

From real clinical encounters to deployed intelligence in 6 steps. Every stage runs on-premises. No patient data ever leaves the clinic.

The 6-Step Pipeline

From Clinical Data to Deployed AI

Step 1

Clinical Data Collection

Real clinical encounters from 8 specialties flow through the NexusOrb EHR: SOAP notes, lab results, prescriptions, vitals, wearable data, nutrition logs, and dream journals. Every data point is timestamped and linked to a patient's 8-axis profile.

Step 2

HIPAA De-Identification

Safe Harbor method removes all 18 HIPAA identifiers. An additional AI NLP scrubbing pass catches embedded PHI in free-text notes. De-identification is verified before any data enters the training pipeline.

Step 3

8-Axis Structuring

De-identified data is mapped to the CH formula's 8 axes: PO, NM, ER, SC, RS, ES, TA, PV. Temporal decay rules model how each axis changes over time, creating longitudinal health trajectories rather than static snapshots.

Step 4

LoRA/QLoRA Fine-Tuning

Llama 3 8B base model is fine-tuned on proprietary 8-axis structured data using LoRA (Low-Rank Adaptation) via MLX on a Mac Studio M5 Ultra. Training runs entirely on-premises. No data is sent to any cloud provider.

Step 5

Clinical Validation

Every model release is benchmarked against standard medical AI tests: MedQA (USMLE), MedMCQA, PubMedQA, and MMLU Medical. Plus proprietary benchmarks for 8-axis cross-reasoning, temporal decay prediction, and drug-outcome correlation.

Step 6

SecureMesh Deployment

Validated models are exported as GGUF files and distributed to clinic Mac Studios via WireGuard VPN (SecureMesh). Inference runs locally at each clinic. No patient queries or responses traverse the public internet.

Privacy by Design

Local-First Architecture

ConceptualHealth.AI is fundamentally different from cloud-based medical AI. No data leaves the clinic. No queries are logged externally. No PHI is ever exposed.

On-Premises Training

Fine-tuning runs on Mac Studio M5 Ultra hardware inside the clinic. MLX framework provides native Apple Silicon optimization. No cloud GPU rentals.

Local Inference

All clinical queries are processed on the clinic's own hardware. Patient questions and AI responses never traverse the public internet.

SecureMesh Updates

Model updates are distributed via WireGuard VPN as GGUF files. Each clinic receives validated model weights through encrypted point-to-point tunnels.

Model Architecture

Purpose-Built for Clinical Reasoning

Multiple model architectures serve different clinical needs, from real-time decision support to complex research analysis.

Llama 3 8B

Active

Phase 1 (Current)

Primary clinical reasoning model. Fine-tuned via LoRA on proprietary 8-axis data. Runs on Mac Studio M5 Ultra with MLX inference.

Llama 3 70B

Planned

Phase 2

Higher-capacity model for complex multi-axis reasoning and long-context clinical narratives. Requires multi-GPU or distributed inference.

Mistral 7B

Planned

Speed Tier

Optimized for low-latency queries: quick lookups, single-axis scoring, and real-time clinical decision support during encounters.

MedGemma

Planned

Imaging

Multimodal medical model for radiology, dermatology, and pathology image analysis integrated with 8-axis patient context.

DeepSeek-R1 671B MoE

Planned

Research

Mixture-of-Experts architecture for complex research queries, population-level analysis, and multi-step clinical reasoning chains.

See It in Action

Request a demo to see ConceptualHealth.AI answer clinical questions that no other medical AI can.

Request a Demo View Benchmarks