LIVESahAIbat DOKIndonesia's Intelligent Scribe for doctorsTry now Learn more
TECHNICAL BRIEF

Indonesia's sovereign medical AI, and the layer that checks it.

Most clinical AI is one model behind an interface. SahAIbat is two planes: a sovereign generative plane that understands, and a deterministic plane that verifies what it produced. The second is why the first can be trusted in a regulated setting — and it is the part competitors would have to rebuild rather than copy.

PLANE 1 · GENERATIVE

Sovereign inference

Indonesia's own medical AI, running on hardware we operate in Jakarta. A medical-specialist base model is being fine-tuned on Indonesian clinical language — how a kader records a danger sign, how a midwife documents ANC 10T, how a doctor writes an assessment in Bahasa Indonesia. Speech, extraction and document reading all resolve on machines under our control, with a governed cloud tier behind them for load.

MedGemma fine-tuneSelf-hosted GPU · JakartaNegation-aware BahasaMultimodal document reading
PLANE 2 · DETERMINISTIC

The verification layer

No model runs here, and that is the point. Clinical arithmetic is computed, never predicted — CKD-EPI, FIB-4, eAG, WHO Z-scores return the same answer every time and can be audited line by line. Diagnosis codes are selected from the ICD-10 catalogue rather than composed, so an invented code is not an error the system is able to make. Evidence is graded strongest-first, and a measured result outranks a mentioned symptom.

Deterministic clinical arithmeticCatalogue-bound codingEvidence-graded reasoningFail-closed by design
AROUND THE MODEL

A model is the easy part. The rest is what a health ministry actually audits.

Risk engine

One field measurement climbs six levels of meaning — from a weight on a scale to a WHO Z-score, a malnutrition class, a referral trigger, a district indicator and a national return, with nobody re-entering it.

Surveillance

Communicable-disease counts aggregate into epidemic curves with alert thresholds derived from the district's own baseline rather than a fixed national number.

Interoperability

SATUSEHAT over HL7 FHIR R4. BPJS via PCare for primary care and E-Klaim for hospital claims, with per-facility credential isolation — nothing pooled, nothing submitted under another clinic's name.

🇮🇩 Data residency

Records are stored in Indonesia on AWS Jakarta, encrypted AES-256-GCM at rest, under UU PDP. SahAIbat is a registered PSE with Kominfo.

Consent-scoped corpus

Corrections become training data only where consent covers it. The corpus is the moat, and it is bounded by the same rules that make it lawful.

Human authority

Every layer is advisory. The system never signs, never submits a claim on its own, and blocks its own output when it cannot justify it.

WHAT THIS PAGE DELIBERATELY DOES NOT SAY

Model weights, training composition, prompt architecture, the verification rules themselves and our inference topology are not published. Everything above describes properties you could confirm from the outside; none of it describes how they are achieved. Under NDA we go considerably further, including a security review and an architecture walkthrough.

REQUEST THE FULL BRIEF

Tell us what you need to evaluate.

We will send the brief that matches the question — a security and residency pack for hospital IT, a model-provenance note for technical diligence, an integration scope for a health office.

Request the brief Usually answered within two working days.