How does self-hosting change the HIPAA / BAA picture?

SynapCores runs entirely inside your environment with local model inference. Embeddings, retrieval, and generation execute where the PHI already lives, so no protected health information transits a third-party AI API. The runtime layer is inside your existing HIPAA scope — there is no new vendor to sign a BAA with for the AI runtime, because there is no external AI vendor in the path. You still control access, encryption, and audit as you do for the rest of the environment.

Does it meet ONC HTI-1 explainability expectations?

The decision support runs on models and retrieval you can inspect. Every suggestion is grounded in cited source records, and every run writes the model, version, inputs, and output to a tamper-evident lineage table — the source-attribute and intervention-transparency posture ONC HTI-1 pushes for predictive decision support. You can show what the algorithm knew and produced at the moment of a recommendation.

Is the AI allowed to act on a patient record?

No — by design. The clinical agent runs with allow_writes=FALSE: it surfaces grounded suggestions and evidence for a clinician, who decides. It informs; it does not act. That human-in-the-loop boundary is enforced in the engine, not left to policy.

Self-hosted clinical AI, inside your HIPAA scope

Clinical decision support and grounded clinical Q&A that run entirely on your infrastructure — the AI runtime lives inside your HIPAA scope, so PHI never leaves and the BAA conversation disappears for the runtime layer. Tamper-evident decision lineage and model versioning for ONC HTI-1.

Book a demo See it run Self-host the free CE

Evaluating for a clinical-informatics team? Self-host it in your environment, then bring us to compliance.

The pain we built this for

If you run clinical informatics or compliance, the AI pitch keeps dying in security review: “The models are promising, but we cannot send PHI to a cloud LLM, we need to show the regulator how a suggestion was made, and we can’t have an AI acting on the chart.”

PHI cannot leave the environment — which disqualifies most cloud-AI decision-support offerings before the pilot.
ONC HTI-1 expects source attributes and intervention transparency for predictive decision support — you need to show what the model knew and cited.
A hallucinated clinical suggestion is a safety event, so ungrounded generation is a non-starter.
An AI must not write to the record autonomously — the clinician has to remain the decision-maker.
Risk models live outside the EHR, so scoring means exporting PHI to a separate ML platform.

See it — a clinical agent that informs but can’t write

An advisory agent bound to cited records, and an encrypted lineage table for the HTI-1 trail — a few statements.

-- A clinical agent that INFORMS but never writes to the record.
CREATE AGENT clinical_advisor
  PERSONA 'Answer only from the cited chart and guideline records. If the
           evidence is not present, say so. Surface options for the
           clinician; never state a definitive diagnosis or order.'
  ALLOW_WRITES FALSE;

-- The ONC HTI-1 lineage: model, version, sources, clinician action.
CREATE IMMUTABLE TABLE cds_lineage (
  encounter_id TEXT,
  model_version TEXT,
  sources      JSONB,          -- cited chart / guideline records
  suggestion   TEXT,
  clinician    TEXT,
  action       TEXT            -- accepted / modified / declined
) WITH (ENCRYPTION = 'AES256GCM');

-- Prove the decision-support trail was never altered.
VERIFY TABLE cds_lineage;

Six capabilities. One self-hosted stack.

Grounded Q&A, in-DB risk models, decision lineage, patient similarity — over one store, inside your HIPAA scope.

Grounded clinical Q&A

Answer clinical questions from your own guidelines, notes, and literature — retrieval-grounded with citations, so every answer traces to a source record instead of model priors.

Vector RAG · citations · EMBED

Deterioration & risk models in-DB

Score readmission risk or personal-baseline vital anomalies directly in SQL — the model runs where the record is, no export of PHI to a separate ML platform.

AutoML · anomaly detection · SQL

Tamper-evident decision lineage

Every suggestion writes model, version, inputs, sources, and clinician action to a hash-chained immutable table. VERIFY TABLE proves the record was never altered — the ONC HTI-1 trail.

CREATE IMMUTABLE TABLE · VERIFY TABLE

Patient similarity & cohorting

Find clinically similar patients with vector similarity plus a relationship graph — cohort discovery and case comparison over one store.

Vector similarity · property graph

Informs, never acts

The clinical agent runs allow_writes=FALSE — it proposes grounded evidence to a clinician who decides. The human-in-the-loop boundary is enforced by the engine.

allow_writes=FALSE · guardrails

PHI never leaves

Local inference inside your environment. No narrative, identifier, or image transits an external AI API — the runtime stays in your HIPAA scope.

Local inference · self-hosted

The objections a compliance lead raises first

PHI & data residency

Self-hosted with local inference — PHI never crosses a network boundary to a cloud model. The AI runtime is inside your existing HIPAA scope.

Explainability & audit (HTI-1)

Grounded, cited suggestions plus tamper-evident lineage of model, version, inputs, and clinician action — reconstructable for review.

Clinical safety

allow_writes=FALSE keeps the agent advisory; a clinician always makes the decision. Grounding makes every claim checkable against the source.

Fits your EHR

Sits beside Epic, Cerner/Oracle Health, or your custom EHR via FHIR/API. It adds grounded, auditable decision support; the EHR stays source of truth.

Run it yourself — the clinical recipe library

Every claim on this page is a runnable recipe — grounded clinical Q&A, decision traceability, risk scoring, patient similarity. Certified against the latest release.

Grounded clinical retrieval

Ground an answer in a citable guideline, self-hosted

Patient similarity & cohorting

Find clinically similar patients by meaning

Tamper-evident CDS lineage (HTI-1)

Model, sources, clinician action — verified

Agentic clinical Q&A

RAG + citations for clinical questions

30-day readmission risk (AutoML)

In-database clinical risk scoring

Browse the full recipe library →

Every recipe is runnable, certified on the latest release

Why one engine beats a stitched pipeline

The RAG that grounds an answer, the model that scores risk, and the lineage that proves it share one store inside your HIPAA scope — no PHI leaves for a separate vector store or cloud model.
The audit trail is complete by construction: source, model, version, and clinician action in one tamper-evident record.
Recipes are SQL you can read, validate, and adapt — not an opaque clinical black box.

5-email course · free

Ship an agent with persistent memory in 5 emails

A 5-email course where you build a real agent with persistent memory — one runnable SQL recipe per email, no LangChain, no Pinecone, no Redis. Lesson 1 hits your inbox the second you sign up. After that: a new recipe every Monday. No spam.

Free forever. We'll never share your email.

Clinical AI your security team can say yes to

PHI never leaves, every suggestion is grounded and logged, and the clinician always decides — self-hosted inside your HIPAA scope. Book a walkthrough, or self-host the free CE.

Book a demo Self-host the free CE Star on GitHub