DIMAGGI AI

The allocation layer of a torus pod: which rectangles a request may take, what a dead chip costs, when isolation and wraparound contradict, and who may reconstitute a slice

Turning scattered GPUs into working AI clusters: capacity accounting, an allocation-policy simulator, control-plane case studies, and a TPU parallel hypothesis. Slurm, Kubernetes, multi-cloud. Reproducible.

Local-first multi-agent deep research orchestration for T3 Code

A machine-checkable standard for AI-cluster chaos engineering: which fault injection tests which layer. Fidelity linter + spec schema + 20-experiment catalog. A pod kill is not an XID; tc/netem is not an InfiniBand flap.

When does network, reliability, or recovery investment beat buying more GPUs? Accounting framework, substitution metric, and decision maps for AI training infrastructure. Paper + reproducible model.

Monte Carlo resilience of the AI-RAN last mile, priced in nines: a single-fiber edge site is ~2.7 nines; multi-transport bonding helps only the tenant classes; the generator — not a third link — buys the decisive nine.

Which GPU-cluster recovery policy wins in which failure regime? A reproducible simulator (ETTR/MTTF/MTTR + dollars) and a policy phase diagram, validated against Meta's published reliability numbers. Spares vs elastic vs checkpointing.

Runtime policy firewall for AI agents

Turning AI infrastructure investment into usable compute: a portfolio covering scheduling, networking, reliability, power, cooling, placement, and governed operations.

Architecture for an autonomous control plane over GPU clusters and networks: separated planes, a referee, an autonomy ladder earned by chaos experiments, and a sourced latency hierarchy — from human intent to nanosecond in-ASIC reflexes.