← Back to Platform
AT Lab · Sovereign Pathology Fabric · v3.13.0

Clinical-grade AI pathology
that never leaves the building.

A federated diagnostic mesh on NVIDIA Blackwell GB10. Hospitals collaboratively train and run multi-specialty pathology models — while raw patient data stays inside the sovereign silo. No cloud. No PHI egress. No vendor lock-in.

🔒 Zero PHI egress 🤝 P2P federated learning 🧬 12+ subspecialties 📜 DICOM WG-26 · FHIR · IEC 62304 ⚡ Blackwell-native FP8
See the GB10 advantage ↓ Why sovereign?
0
tests · 0 failures
0
clinical subspecialties
0
% data stays on-prem
0
SaaS dependencies
The impasse

Cloud pathology asks you to ship the patient out.

Gigapixel slides + PHI in someone else's data center means compliance risk, recurring egress cost, and a model you don't control. The Sovereign Fabric inverts that.

Cloud SaaS pathology

🌐 Your data, their building

Slides + PHI uploaded to a multi-tenant cloud. Per-study egress fees. A black-box model that updates without your sign-off. Network outage = no diagnosis. Audit trail you can't fully see.

AT Lab Sovereign Fabric

🏛️ Your data, your building

Inference + training run on the local Blackwell node. Only encrypted gradient slices cross the Tailscale mesh — never pixels. TPM-signed reports, hash-chained model lineage, and a deterministic safety gate you control. Works fully offline.

One platform

Depth across the whole pathology matrix.

From breast oncology morphometrics to the clinical-pathology pillar the image models can't serve — anatomic, clinical, molecular, and beyond.

🏅

Nottingham grading

Tubule + pleomorphism + mitoses assembled into Grade 1–3, pre-graded during background scan.

🛡️

TME immune phenotype

Per-nucleus phenotyping → infiltrated / excluded / desert from TIL–tumour spatial enrichment.

🧫

StarDist + Cellpose

Star-convex nuclei in dense fields, IHC H-score / HER2 / Ki-67 hotspot auto-selection.

🗺️

Whole-slide tissue map

Tumour / stroma / necrosis composition + colorized overlay, cached at scan time.

🔗

Serial-section registration

FFT phase-correlation aligns H&E ↔ IHC so annotations transfer across stains.

📐

Spatial statistics

Clark-Evans & Ripley's K — clustered / random / dispersed cell architecture.

🧬

Molecular / NGS

Gene · alteration · VAF · AMP tier · MSI / TMB extracted from the narrative.

🦠

Microbiology

Organism ID + antibiogram (S/I/R) with MDR flagging.

🩸

Transfusion & heme

ABO/Rh, DAT, antibody panel, parasitemia quantification.

🫘

Banff transplant

Full renal-allograft lesion scoring → ABMR / borderline / TCMR classification.

❤️

Cardiac & valve

ISHLT rejection grade + valve scoring (calcific / myxomatous / endocarditis).

⚖️

Forensic autopsy

Cause-of-death chain, manner classification, organ weights, blood-alcohol.

🎯

Attention-MIL slide call

Per-tile scores pooled into a whole-slide probability + attention heatmap.

📊

Conformal uncertainty

Calibrated abstention — low-confidence cases routed to human review automatically.

🚦

Slide QC gate

Focus / artifact triage (pass / review / fail) before inference even runs.

Multi-specialty routing

One shared Blackwell pool serves pathology, radiology, dermatology, odontology.

🔁

Longitudinal tracking

Progression / regression vs a patient's priors — grade, TIL, immune phenotype deltas.

🗂️

Cohort & worklist

Auto-prioritized worklist + cohort analytics across signed cases.

🔐

Confidential computing

In-VRAM TEE attestation — a peer verifies the enclave before trusting gradients.

⛓️

Model-lineage ledger

Tamper-evident hash chain: diagnosis → model card → validation verdict → attestation.

🛂

Deterministic safety gate

Every consensus adapter is regression-tested on a golden cohort — auto-rejected on drop.

📈

Drift monitoring

Per-task AUC EWMA + trend → warning / critical alerts before the hard gate trips.

🏥

Standards-native

DICOM WG-26 (QIDO/WADO/STOW/ANN), HL7 ORU, FHIR R4, CAP eCC / NAACCR export.

🧾

IEC 62304 traceability

CI emits a CycloneDX SBOM + software-unit → test matrix for SaMD audit.

The Blackwell GB10 advantage · interactive

One memory pool. Engineered to never starve.

GB10 shares a single physical memory pool between CPU and GPU. Naïve vLLM settings reserve 90 %+ of the whole node for idle KV cache. Drag the controls — watch the Sovereign Fabric's right-sizing + idle-reaper reclaim the node in real time.

Blackwell-native · ~½ weight memory
Lazy-start on demand
Node: NVIDIA GB10 · 128 GB unified memory
% used
27B 4B GigaPath CPU services free headroom

Illustrative model of the documented tuning in docker-compose.yml + engine_reaper.py. Default settings reclaim ~50 GB vs. the naïve configuration — with zero change to diagnostic output.

Why it wins the deal

The economics of sovereignty.

💸

$0 egress

No per-study upload fees. Slides never leave the NVMe array next to the GPU.

🔌

Works offline

A network outage doesn't stop a single diagnosis. The silo is self-sufficient.

🤝

Collective intelligence

Every site's model improves from the network — without any site sharing data.

🛡️

Audit-ready by design

TPM seals, lineage chain, SBOM, traceability matrix — the SaMD paperwork is generated, not bolted on.

🚀

One box, many specialties

Memory-smart residency means a single GB10 serves pathology + radiology + derm + dental.

🔓

No lock-in

Open standards in, open standards out. Your models, your weights, your registry.