Infrastructure software

DIMAGGI builds AI infrastructure software for compute, networks, workload placement, reliability and control.

Connected infrastructure

Compute, network paths, placement and recovery constrain one another. Our tools make those relationships explicit.

Compute architecture and capacity: Compare workload demand with accelerator capacity and resource shape.
Network and interconnect: Check communication requirements against topology, locality and path constraints.
Workload orchestration and placement: Check where a workload can run. Workflow coordination and resource scheduling have different roles.
Reliability and recovery: Compare retained work, checkpoint costs and recovery choices.
Infrastructure control: Check proposed actions against policy, system state and approval requirements.

Power, cooling, facility space and site topology are inputs to these decisions.

Agentic engineering

We use software agents to implement changes and run checks. Engineers review the changes before integration.

Define requirements
Implement
Test affected interfaces
Review
Integrate

Example workflow: an agent updates a Tool Guard policy rule, runs policy tests, and prepares the change for review. This is a development workflow, not permission to operate infrastructure.

Infrastructure control example

Recorded validator execution on synthetic inputs.

SPAN

Not checked

XP1 declared-vs-computed blast radius: no plant graph supplied, so the envelope's blast_radius was taken on trust
XP2 topology currency: no live topology hash supplied, so the envelope's topology_hash was taken on trust
TN1 organization slice quota: no tenancy block on the envelope, so the organization's slice count in 'hall-a' was taken on trust
TN2-TN4 wavelength sharing: no tenancy block on the envelope, so whatever else is on the stitch's wavelength was taken on trust

These are admission decisions, not executed infrastructure actions.

Inputs, output and reproduction

Exit code: 0

PYTHONPATH=src python3 -m spancontract.cli validate examples/campus-stitch-pipeline.json --json

{
  "envelope": {
    "scale_out": "campus_fiber",
    "span_mode": "span",
    "span_rtt_us": 42.0,
    "span_bw_gbps": 800.0,
    "span_failure_domain": "campus-duct-east",
    "slice_rect": {
      "hall_id": "hall-a",
      "origin": [
        0,
        0,
        0
      ],
      "extent": [
        8,
        8,
        8
      ]
    },
    "stitch_id": "stitch-ab",
    "stitch_api": "https://ocs.plant.example/v1",
    "topology_hash": "b7c2000000000000000000000000000000000000000000000000000000000000",
    "measured_il_db": 6.0,
    "measured_ber": 1e-12,
    "checkpoint_window_s": 120.0,
    "collective_window_s": 300.0,
    "ingest_budget_GBps": 40.0,
    "thermal_headroom_k": 3.0,
    "ride_through_s": 180.0,
    "power_headroom_kw": 250.0,
    "compile_cache_key": "aeb84a2c41d76cea9b6e500f38a2624cf780de86343ebe7ada6b8c791b03206d",
    "blast_radius": 1,
    "autonomy_level": "L1",
    "requested_action": "train",
    "measured_age_s": 30.0,
    "stitch_api_reachable": true,
    "labels": {
      "graph_hash": "a3f1000000000000000000000000000000000000000000000000000000000000",
      "cut": "pp"
    },
    "tenancy": null
  },
  "decision": "span",
  "findings": [],
  "not_checked": [
    "XP1 declared-vs-computed blast radius: no plant graph supplied, so the envelope's blast_radius was taken on trust",
    "XP2 topology currency: no live topology hash supplied, so the envelope's topology_hash was taken on trust",
    "TN1 organization slice quota: no tenancy block on the envelope, so the organization's slice count in 'hall-a' was taken on trust",
    "TN2-TN4 wavelength sharing: no tenancy block on the envelope, so whatever else is on the stitch's wavelength was taken on trust"
  ],
  "policy": {
    "measurement_ttl_s": 300.0,
    "sync_training_max_rtt_us": 2000.0,
    "tensor_parallel_max_rtt_us": 0.0,
    "max_insertion_loss_db": 20.0,
    "max_bit_error_rate": 1e-09,
    "max_autonomous_blast_radius": 1,
    "training_autonomy_levels": [
      "L0",
      "L1"
    ],
    "training_actions": [
      "train",
      "pretrain",
      "finetune",
      "continue_pretrain"
    ],
    "require_checkpoint_within_ride_through": true,
    "min_thermal_headroom_k": 0.0,
    "min_power_headroom_kw": 0.0,
    "enforce_org_slice_quota": true,
    "allow_lambda_sharing_across_orgs": false,
    "dedicated_tenancy_classes": [
      "dedicated"
    ]
  },
  "chain": "4c24e46239eab54786f8dbec240ccf35f0720ce5434bb6781e41e190585327e4"
}

Tool Guard

Policy checks before action. An audit record after every decision.