A privacy-first framework for provenance-aware multimodal memory and orchestration in high-stakes AI workflows.

vai_sante_os is an exploratory research infrastructure project. It is not a clinical product, not a medical device, and not a decision-maker. The goal is to make it easier to prototype and evaluate AI systems that need to reason over mixed evidence safely, with traceability and human oversight.

Most AI demos look good on happy-path prompts and fall apart when you ask:

• Where did this claim come from?
• Can I inspect the original evidence?
• What changed over time?
• How do I review or override this safely?

This project focuses on those harder questions.

High-stakes workflows (health, legal, policy, safety, operations) often involve:

• multimodal evidence (text, scans, audio, forms, structured records)
• longitudinal context (state changes over weeks or years)
• strict provenance expectations
• privacy constraints and local-first data handling
• human review at critical decisions

vai_sante_os explores a generalisable architecture for that setting.

• Privacy-first multimodal memory
  Keep storage boundaries explicit and minimise data movement.

• Provenance-aware retrieval
  Retrieval returns both content and its chain of custody.

• Longitudinal reasoning scaffolds
  Treat time as first-class, not as stray metadata.

• Human-in-the-loop orchestration
  Route sensitive steps through review gates.

• Evaluation and governance by default
  Assess uncertainty, evidence quality, and failure modes early.

flowchart TD
    A[Ingestion Adapters\nPDF, OCR, Audio, Structured Data] --> B[Normalisation & Redaction]
    B --> C[Evidence Store\nObject + Metadata + Access Policy]
    C --> D[Indexing Layer\nEmbeddings, Entity Graph, Temporal Index]
    D --> E[Retrieval Engine\nHybrid + Provenance Constraints]
    E --> F[Reasoning & Orchestration\nAgent Workflows + Tools]
    F --> G[Review Interface\nHuman Approval + Feedback]
    G --> H[Audit & Evaluation\nTrace Logs + Metrics + Replay]

sequenceDiagram
    participant S as Source
    participant I as Ingestion
    participant M as Memory
    participant R as Retrieval
    participant O as Orchestrator
    participant H as Human Reviewer

    S->>I: Raw multimodal input
    I->>M: Normalised artefacts + provenance metadata
    O->>R: Query with policy + task intent
    R->>O: Ranked evidence + citations + confidence cues
    O->>H: Draft output + rationale + risk flags
    H->>O: Approve / request revision / reject
    O->>M: Store decision trace + feedback signals

.
├── README.md
├── LICENSE
├── .gitignore
├── docs/
│   ├── architecture.md
│   ├── evaluation.md
│   ├── governance.md
│   ├── threat-model.md
│   ├── roadmap.md
│   └── status.md
└── templates/
    ├── eval-scorecard.md
    └── incident-review.md

• No “black box says so” outputs in high-stakes settings.
• Every material claim should be back-trackable to source evidence.
• Confidence is not enough; uncertainty and missingness should be explicit.
• Human escalation should be cheap and normal, not treated as failure.
• Offline replay should be possible for audits and post-incident learning.

See:

• docs/evaluation.md
• docs/threat-model.md
• docs/governance.md

This repository is currently a research showcase scaffold:

• architecture and operating model documented
• safety, evaluation, and governance artefacts stubbed
• implementation modules intentionally lightweight

It is suitable for:

• design reviews
• portfolio discussion
• exploratory prototyping

It is not suitable for direct real-world deployment without substantial engineering, security hardening, and domain validation.

Historically, this architecture started from a healthcare/oncology context. In this repository, it is framed as a general high-stakes evidence workflow framework.

1. Read docs/status.md
2. Read docs/architecture.md
3. Review risk controls in docs/threat-model.md
4. Use templates under templates/ for evaluation and incident reviews

Apache License 2.0. See LICENSE.