Operator-grade advisory for AI infrastructure.

Situation

A stealth neocloud needed to bring its first H100 cluster online to hit a contractual revenue milestone — 50 staff across 10 teams, no shared deployment playbook, no production track record.

Approach

Built the first baremetal deployment workflow end-to-end. Stood up daily node-tracking, defined exit criteria for inter-team handoffs, and routed every gate through a single Jira system of record.

Outcome

First revenue milestone delivered on schedule. Over 2,000 production changes tracked through the system. The deployment workflow became the operating standard across subsequent cluster classes.

Situation

An AI infrastructure provider needed to build a competitive multi-tenant storage offering against FsX Lustre and S3, with a real customer waiting on capacity decisions.

Approach

Ran vendor diligence across Weka, Vast, Alluxio, Quobyte, Tigris, Pure, Cloudflare R2, Solidigm, WD. Designed a two-tier architecture — high-performance parallel filesystem plus on-prem object — and negotiated preferential pricing.

Outcome

Closed the deal in weeks at healthy margin. First service-availability metrics shipped for S3, POSIX, and control-plane APIs. MLPerf Storage benchmarks established the public baseline.

Situation

Engineering teams were flying blind. Node failures surfaced through customer escalations, not telemetry. Inventory reconciliation was manual across three sources of truth.

Approach

Stood up a Grafana-fronted telemetry stack. Unified inventory through a ServiceNow + Jira stack, Netbox, and a Snowflake source of record. Wired automated health checks into a hot-spare and remediation loop.

Outcome

Mean-time-to-detect dropped from hours to minutes. Automated failover became routine. Inventory reconciliation moved from a weekly meeting to a passive dashboard.

Situation

A hyperscaler partner was scoping a sovereign cloud buildout in the 1 GW range, with 800V DC architecture and a non-trivial transition from baremetal to virtualized GPU offerings.

Approach

Reviewed the 800V DC spec, baremetal automation, and firmware-detection layer. Built a technical due-diligence template applied across site evaluations from XX MW to X GW. Scoped backbone connectivity at multi-Tbps PNI.

Outcome

Validated the BMaaS-to-VMaaS migration path for the chip OEM. Defined networking WAN scope across multiple stakeholders. Template now reusable across subsequent sovereign engagements.

Situation

A stealth AI cloud needed its first customer-facing network: a multi-tenant overlay, load balancers, and management plane purpose-built for AI training and inference — not a re-skin of enterprise cloud networking.

Approach

Architected the customer overlay end-to-end and directly contributed control-plane and datapath code. Partnered with AI server and Kubernetes engineers on on-host networking and NCCL, and led cross-layer triage when issues spanned GPU hosts and the underlying fabric.

Outcome

Overlay control plane and datapath running in production in the first AI datacenter. Design decisions across the network software stack realigned to AI workload demand.