On-premise medical inference on the GB10 workstation
An NVIDIA GB10 workstation with 128 GB of unified memory holds models up to 200B parameters under a desk. For health data the constraint is not capacity but memory bandwidth: 273 GB/s.
Read moreBlog
Articles on AI governance, cybersecurity and systems architecture.
2026
Trustworthy AI for Healthcare Systems
Healthcare demands AI systems that are verifiable, explainable, and compliant with rigorous clinical and regulatory standards.
Read more2025
Running an LLM over clinical data without it leaving the hospital
A language model in contact with records and reports has to stay inside the perimeter: PII redaction upstream, immutable logging for the EU AI Act, risk classification under MDR. The architectural constraints that follow.
Read more
CDA, FHIR and openEHR: three standards, three layers of a health system
CDA R2, FHIR and openEHR do not sit on the same plane of a health system: they answer three different questions — archiving, exchange, keeping data over time. The roles, the contact points, and the friction that appears when a standard ends up in the wrong place.
Read more
EHDS: Regulation (EU) 2025/327 on European health data
Regulation (EU) 2025/327 on the European Health Data Space has been in force since 26 March 2025. A technical reading of primary use, secondary use, the Health Data Access Body and the application timeline.
Read more
Where the open part ends in NVIDIA's DGX stack
In NVIDIA's DGX platforms the interconnect silicon (NVLink, NVSwitch) is proprietary, while the software and mechanical layers around it are public: NCCL under a BSD licence, the GB200 NVL72 rack contributed to the Open Compute Project. Where the line falls between what is documented and reusable and what stays tied to a single vendor.
Read more
Open weights and licences: what DeepSeek-R1 actually releases
DeepSeek-R1 ships with weights under the MIT licence, but the open weights label covers the parameters alone: data, training code and the RL procedure stay out. What that changes for anyone who has to audit the model.
Read more
Recognising an agent before you can stop it
Governing an autonomous agent needs three primitives: identification, diagnosis, intervention. They sit at very different levels of maturity, and the asymmetry is the main operational constraint.
Read more2024
The national FSE 2.0 Gateway and the 7 September 2023 decree
The Italian Ministry of Health decree of 7 September 2023 (Official Gazette no. 249 of 24 October 2023) introduces the Sogei-run Gateway and the regional Healthcare Data Ecosystem. Document formats, IHE profiles and the FSE 2.0 feeding model as of September 2024.
Read more
An LLM on a Raspberry Pi 5: local inference with no network
With the 1-4 billion parameter models released in April 2024 and GGUF quantisation, a Raspberry Pi 5 runs an LLM with no connection. That is a constraint of architecture, not a performance figure: text that never leaves the device is subject to no transfer at all.
Read more
How open is the open-source NVIDIA GPU driver
NVIDIA's Linux GPU kernel modules have been open source since 2022, but the boundary of that openness runs through the GSP and the closed CUDA user space. What stays proprietary, and why the open driver depends on a closed firmware.
Read more
IPS and MyHealth@EU: the content standard and the exchange infrastructure
The International Patient Summary (EN ISO 27269) defines what a patient summary contains; MyHealth@EU how it travels between Member States. Why keep the two layers apart, and why semantic mapping stays the fragile part.
Read more
Neural networks and quantum annealing on power grids
Applying graph neural networks and quantum annealing to grid control opens a verifiability problem that accuracy metrics do not cover.
Read more
TinyLlama: a training pipeline third parties can reconstruct
TinyLlama 1.1B releases weights, code, datasets and intermediate checkpoints under Apache 2.0. For governance the value lies in the chain from data to weights, inspectable step by step, regardless of model size.
Read more2023
When an Arduino sensor becomes a medical device
The regulatory line between an ESP32 + MQTT node carrying a MAX30102 sensor and a medical device under Regulation (EU) 2017/745. What changes between a telemonitoring demo and something you can put in a patient's hands.
Read more
Security Onion 2.4: one Elastic Agent in place of three endpoint agents
With the 2.4 series Security Onion collects host data from a single Fleet-managed Elastic Agent, in place of Wazuh, Beats and osquery. Technical notes on architecture, Salt orchestration and licensing.
Read more
What VALKYRIES leaves behind: interoperability in cross-border disasters
The European VALKYRIES project closes in September 2023. The recurring technical problem is semantic: emergency services from different countries lack a shared vocabulary, before they lack the technology to exchange it.
Read more
Integrating a diagnostic model into a laboratory automation system
Placing a deep-learning prediction into the data flow of a Total Laboratory Automation system imposes constraints on protocols, latency and traceability that come before the quality of the model.
Read more
MONAI Label: assisted annotation for 3D imaging
MONAI Label moves the cost of medical annotation from manual drawing to correcting a proposal: a REST server, plugins for 3D Slicer and OHIF, and active learning driven by model uncertainty.
Read more
SQLAlchemy 2.0: typing moves into the ORM mapping
With 2.0 the declarative layer derives columns from PEP 484 annotations and the select() API becomes the single form for Core and ORM. What actually changes for anyone maintaining a Python application.
Read more