Collaborating with industry leaders
HIRO PEaaS STACK
Dataset and model versioning
Keep datasets, models and training history organized, with each model version linked to the dataset and parameters used in training.
Reusable training workflows
Improve models through repeatable experiments: adapt existing workflows and compare parameters, metrics and results to see what works best.
Hardware-aware deployments
Serve models to your applications on your own infrastructure, with compute and memory sized to each model’s requirements.
Controlled model updates
Shift traffic to a new classical model revision in 5% steps on KServe, and roll back without redeploying.
Shared data catalog
Find training data across biobanks and partner organizations, with clear usage terms and descriptions.
Owner-defined access
Control who can find datasets and start federated training by role and organization. Issue time-limited digital contracts and revoke them when needed.
Federated learning
Train models and run analysis on data that partners cannot transfer. Approved workloads run at source, under the owner’s rules.
Signed contracts and contract log
Cover each training order with a digitally signed contract that expires or can be revoked, recorded in each site’s contract log.
Autonomous sites
Each site operates on its own and coordinates with other sites peer to peer. If one site fails, the others take over its applications.
Adaptive scheduling
Maintain application performance by moving workloads away from overloaded or unhealthy nodes.
Automatic rightsizing
Free up capacity for other applications by adjusting resource allocations to actual usage.
Predictive autoscaling
Maintain performance and uptime by scaling capacity before demand spikes.
Sized to the site
Hang a unit on an office wall, or stack several units in a server closet. Add more units if the site needs more compute.
Flexible configuration
Fill the unit with the GPUs, CPUs and storage your workload needs. Use server modules from any vendor that builds to the open COM-HPC standard.
Low power use
A fully loaded unit draws under 3 kW. Plug it into an ordinary 16 A circuit, with no upgrade to the site’s power supply.
Fans or water
Cool the unit with fans, or connect it to the building’s water loop.

NextGen
Training AI on hospital data without exposing patient records

GLACIATION
Distributing data processing for Italy’s national payroll service

ACES
Extending power-grid analytics across distributed edge infrastructure

CAPE
Building edge infrastructure that can evolve without a full rebuild

Shift2DC
Reducing power losses in edge data centers with direct current

MYRTUS
Keeping distributed data consistent from edge to cloud

SYCLOPS
Keeping AI acceleration open and portable across hardware platforms

BRAINE
Bringing data-center capacity to the network edge
OptimAIse
Running generative AI efficiently on European hardware

VERGE
Running secure, explainable AI on 5G edge infrastructure






