AI datacenter fabric512-GPU training fabric, lossless from day one
A model lab in Texas needed a back-end fabric for 64 GPU nodes that would not stall collectives. We designed a rail-optimized 400G RoCEv2 fabric on Arista 7060X, tuned PFC/ECN and buffer profiles on the twin, and benchmarked before handover.
Scope · 64 nodes, 512 GPUs, 8 rails, 2 spines, 16 leaves
Result · NCCL all-reduce at 97% of theoretical on handover
Delivered · fabric as code, twin, live portal
Timeline · 11 weeks from design freeze to first training job
Cisco ACI to Juniper Apstra without a maintenance weekend
An enterprise datacenter with 38 leaves and 400+ EPG contracts wanted multi-vendor freedom. Contracts were mapped to Apstra blueprints and security zones, each wave rehearsed on the twin, and legacy decommissioned only after verification.
Scope · 38 leaves, 4 spines, 2 pods, 412 contracts
Result · zero rollbacks used across six waves
Delivered · blueprint as code, repository, runbooks
Timeline · one quarter, all cutovers in normal windows
ASA and FortiGate estate consolidated onto Palo Alto
A regional financial institution ran three firewall platforms with 1,400 rules nobody dared remove. Policy was translated, de-duplicated and tested against captured production flows before three waves of cutover.
Scope · 14 firewalls, 1,400 rules, 3 platforms
Result · zero rule loss, 31% fewer rules, Panorama as single source
Delivered · policy as code, twin test harness, evidence for the regulator
Timeline · 9 weeks
Juniper Mist campus for a 14-site healthcare group
Legacy Cisco campuses at 14 clinics and a hospital moved to an EVPN campus fabric on Juniper EX with Mist, Wi-Fi 7 and medical-device segmentation, operated under our AIOps portal.
Scope · 14 sites, 212 switches, 640 APs
Result · ticket volume down 40% in the first quarter
Delivered · site-by-site cutover, Marvis run for the customer
Timeline · two quarters, academic-style calendar constraints respected
MPLS exit to Fortinet Secure SD-WAN with Alkira cloud exchange
A distributed manufacturer with 60 branches backhauled every cloud session through one datacenter. Branches moved to FortiGate SD-WAN in waves; AWS, Azure and SaaS entered through Alkira with inspection applied once.
Scope · 60 branches, 2 datacenters, 3 clouds
Result · MPLS cost down 58%, cloud latency down 70%
Delivered · templates as code, SLA dashboard in the portal
Timeline · 5 months, no flag-day cutover
Twin-gated change pipeline for a colocation provider
A multi-tenant operator with weekly fabric changes and a history of outages adopted our twin-and-repository workflow. Every change now passes a Batfish and Containerlab run before the window.
Scope · 3 fabrics, 120 switches, ~40 changes a month
Result · change-related incidents to zero over two quarters
Delivered · pipeline, twin, drift detection, training
Timeline · 6 weeks to first gated change
