Abdoul A.

Group-wide cyber resilience program protecting ~1,000 vital applications under ECB requirements for systemic institutions.

• • Secured continuity of ~1,000 vital banking applications under ECB and DORA requirements, holding RTO <10 min / RPO 30 min across a 45+ day isolation mandate, by architecting an air-gapped EVPN/VXLAN fabric on Arista CloudVision.

• • Enforced full east-west isolation through micro-segmentation, a unidirectional data-diode ingestion path, and a Zero Trust access chain (CyberArk PSM, RSA MFA, Azure AVD); isolation patterns directly transferable to secure multi-tenant AI workloads.

• • Established the platform's observability and reliability framework (SLO definition, error budgets, blameless post-mortems) on Zabbix and Nagios monitoring.

• • Drove architecture decisions across security, infrastructure, and operations stakeholders, delivering the SRD/HLD/LLD/DAT and runbooks for the program.

Technologies: Arista CloudVision

◦ · Data diode

◦ · Zabbix

◦ · Nagios

Full-stack GPU training and inference platform: end-to-end architecture across network, compute, storage, and operations. Client and details under NDA; references available on request.

• • Scaled an HPC-class GPU training/inference fabric from a 64-GPU pilot to a 2,048+ GPU multi-pod design, enabling 70B-parameter model training and 10k+ QPS inference, with engineered headroom for 4x scale-out and no redesign.

• • Engineered a rail-optimized RoCEv2 Clos fabric with ECMP flow-let switching and tuned PFC/ECN/DCQCN congestion control, designed to sustain >90% of theoretical AllReduce bandwidth with <2% degradation under single-switch failure.

• • Removed checkpoint I/O as a training bottleneck, reaching >25 G B/s checkpoint throughput per server, by engineering an NVMe-oF over RoCE storage fabric; isolated backend, storage, frontend inference, and OOB in a multi-fabric, zero-trust design.

• • Made fabric provisioning repeatable and error-resistant by codifying the full underlay/overlay as Infrastructure as-Code (Terraform, Ansible) with GitLab CI/CD validation on Linux-based tooling.

• • Owned architecture decisions across a cross-functional program from strategy through production, delivering the BOM, validation plan, and operational model.

Technologies: NVIDIA ConnectX-7

◦ · RoCEv2 / RDMA

◦ · EVPN/VXLAN

◦ · PFC/ECN/DCQCN

◦ · NCCL

◦ · NDFC

◦ · PTP

◦ · NVMe oF

◦ · Terraform

◦ · OpenTofu

◦ · Ansible

◦ · Python

◦ · GitLab CI/CD

◦ · Linux

Ground-up infrastructure reconstruction for the transfer of HSBC France retail banking operations (1M+ customers) to My Money Bank.

• • Delivered the ground-up datacenter network migrating 1M+ HSBC France retail customers to a new bank, hitting a non-negotiable regulatory cutover with zero client impact and zero operational disruption.

• • Built a production EVPN/VXLAN spine-leaf fabric on Arista EOS (BGP EVPN, ECMP, fast convergence) for high-performance east-west traffic, with the full HLD/LLD, design standards, and BOM.

• • Standardized provisioning through Arista CloudVision (CVP) automation and Bash/Python tooling on Linux, reducing manual configuration errors and making rollouts repeatable.

• • Layered VMware NSX micro-segmentation over an active-passive dual-DC design, defined AWS connectivity via Direct Connect, and delivered branch LAN, Cisco Meraki SD-WAN, and perimeter security (Palo Alto, F5).

Technologies: Arista EOS

◦ · CloudVision Portal (CVP)

◦ · EVPN/VXLAN

◦ · BGP EVPN

◦ · ECMP

◦ · VMware NSX

◦ · Cisco Meraki

◦ · Fortinet SD-WAN

◦ · AWS (Direct Connect)

◦ · Palo Alto Networks

◦ · F5

◦ · Bash

◦ · Linux