GPU and CPU compute, Linux and scientific software, data pipelines, parallel storage and low-latency networking, engineered as one system. Two decades designing, building and operating on-premises and cloud infrastructure for AI training, big-data workloads and scientific computing. We build and lead engineering teams, and help grow the communities around AI and HPC centers: researchers, platform engineers and operators working together. From a single GPU server to AI factories and research centers, our work spans architecture, deployment, performance tuning and day-to-day operation.
Compute, fabric and storage are one system, and the failures that matter happen between them rather than inside any one of them. We work across all three, from architecture and procurement through build-out, benchmarking and day-to-day operation.
Full stack design and delivery: hardware selection, provisioning, operating system and image management, node classes, and the capacity planning behind them. Bright Cluster Manager, OpenHPC, HPCM, xCAT. Multi-site and multi-region estates.
Fabric design and management from EDR through HDR and NDR, fat tree and dragonfly topologies, OpenSM routing engines, NVIDIA UFM, RoCE and RDMA. Switch and adapter firmware, port configuration, and the diagnostics that separate a fabric problem from an application problem.
GPFS and IBM Storage Scale, Lustre, WekaFS, VAST, Pure, Isilon, DDN, BeeGFS and CEPH. Tiering and policy, erasure coding, data movement, and benchmarking that measures the workload rather than the page cache.
A100, H100, DGX systems, NVLink, DCGM, Intel Max and AMD Instinct. CUDA and ROCm, NCCL collective tuning, PyTorch and TensorFlow environments, containerized delivery through Apptainer and Kubernetes.
Slurm, PBS Pro, LSF, Grid Engine and Univa. Partition and QoS design, GPU and license-aware scheduling, fair share, accounting, and the migrations between schedulers that organizations put off for years.
AWS ParallelCluster, SOCA and IDEA, FSx for Lustre, EFA and the surrounding services. Azure HPC across regions. Lift and shift from on-premises, and the more common case where the honest answer is that only part of the workload should move.
HPC does not sit on its own. It depends on identity, network, storage, monitoring and compliance that have to work, and on people who can run them. AI is changing how each of these is done, and the shift is mostly practical rather than dramatic.
Linux and UNIX estates at scale, identity and directory services, virtualization, configuration management with Ansible and Puppet, monitoring through Prometheus, Grafana, Nagios and Ganglia. Migrations off legacy platforms, including Solaris and HP-UX, where the hard part is the application stack rather than the operating system.
Where AI changes it: telemetry has always been collected faster than anyone could read it. Anomaly detection across metrics and logs turns that backlog into something that surfaces problems before a user reports them, and capacity forecasting stops being a spreadsheet exercise.
Spine and leaf design, 10, 25 and 100 gigabit platforms, Juniper, Arista, Cisco and Brocade, OSPF and BGP, VXLAN and MLAG. Campus and transcontinental network optimization for research and engineering traffic that behaves nothing like office traffic.
Where AI changes it: traffic classification and fault correlation across a fabric this size is pattern work, and pattern work is where machine learning earns its place. It shortens the gap between a link degrading and someone knowing which link.
Building and running research and clinical computing under the regimes that govern it: CUI handling to NIST SP 800-171, NIH controlled-access genomic data, HIPAA, GxP and PCI. SELinux and RBAC, vulnerability management, access and filesystem auditing, threat management across cloud and on-premises. NACI Public Trust clearance. Security officer responsibility on a federal modeling platform.
Where AI changes it: evidence gathering. Most audit effort goes into collecting and correlating what a control actually did, not into judgment. Automating that collection leaves more room for the part that needs a person.
Petascale storage planning, purchase and migration. Backup and archive across enterprise and cluster filesystems, tape libraries, snapshot and replication frameworks. Transfer at distance with Aspera and GridFTP. Retention and access control in environments where the data has legal weight.
Where AI changes it: placement. Deciding what belongs on flash, what belongs on capacity tiers and what should have been deleted is a prediction problem, and policy engines have historically guessed with very blunt rules.
Group lead and IT director roles: hiring and technical screening, mentorship, team structure, vendor and contract management, budgeting and expansion planning. Running engineering, development and support functions together rather than as separate silos.
Where AI changes it: mostly in what a small team can now cover. The practical question is which work it removes and which work it quietly creates, and that is worth answering before rather than after a reorganization.
Research computing is rarely unregulated. Genomic data arrives with a data use agreement attached, a DoD subcontract brings CUI obligations onto your storage, and the moment a foreign national gets a login on a machine holding controlled technology you are in export control whether anyone planned for it or not. These are the frameworks the work has been delivered under — described as engagements, not as badges.
NIST SP 800-171 across storage, scheduler and identity, with the system security plan and POA&M that go with it. NIST SP 800-53 on federal systems, FISMA responsibility as a named security officer, and CMMC readiness for defense subcontracts. DISA STIG and CIS Benchmark hardening on compute at scale. NIST SP 800-88 media sanitization on decommission. FIPS 140 validated cryptography where it is required, and the exemption argument where a validated module does not exist for the hardware.
The NIH Genomic Data Sharing Policy and the dbGaP security best practices as they actually apply to a shared cluster: where controlled-access data may live, who may reach it, how access is logged, and how it is destroyed at the end of the approved period. NIH Data Management and Sharing plans, data use agreements and honest broker arrangements, IRB-driven access boundaries, and the awkward part — enforcing per-study separation on a filesystem several hundred other users also mount.
EAR and ECCN classification of HPC hardware, GPUs and the software stack, and ITAR where it applies. Deemed export is the one that catches universities: giving a foreign national an account on a machine holding controlled technology is an export, and shared clusters are built to do exactly that. Technology control plans, restricted party and OFAC screening, Entity List constraints on accelerator procurement, and where the fundamental research exclusion genuinely applies rather than where people wish it did.
HIPAA Security Rule on estates that are simultaneously research and clinical, GxP and 21 CFR Part 11 for electronic records and signatures, PCI DSS, and GDPR where a collaboration crosses into the EU. Audit evidence that can be produced on demand rather than reconstructed the week before an assessment.
Systems designed and operated to ISO/IEC 27001 information security controls, ISO/IEC 27017 and 27018 for cloud and personal data in cloud, ISO/IEC 20000 service management practice, and SOC 2 control objectives. Where an organization is pursuing certification, the work is the control implementation and the evidence that survives the audit.
Compliance fails on shared HPC in predictable places: a scratch filesystem nobody scoped, an epilog that never deletes, a scheduler prolog that leaks a path, home directories mounted everywhere by default, node images drifting apart, and audit logging that stops at the login node. The useful work is closing those, then proving they are closed — not producing a policy document that describes a system nobody has actually configured.
Roles as HPC architect, engineer, group lead and IT director, in environments where the machines are expensive and the deadlines are external.
NVIDIA AI factory environments, GPU fabric, scheduler and parallel storage integration
Node liquid-cooled Cray and HPE clusters, federal weather modeling
Nodes across three sites in two continents, semiconductor manufacturing
Nodes migrated from on-premises Linux and HP-UX to cloud, automotive
Users on a mixed research and clinical estate, academic medicine
High-throughput HPC estates in academic medicine, clinical research and defense postgraduate education
Those numbers are where the work has led. They are not the price of admission, and they are not the minimum engagement. A single GPU server that is not delivering the throughput it should, a storage mount that keeps dropping under load, a scheduler nobody has retuned since the day it was installed, a benchmark that disagrees with the vendor's datasheet — these are the same problems as the large ones with fewer zeros on the end, and most of them are an afternoon or a week rather than a project. If that is all you need, that is what we will quote, and if you do not need outside help at all we will tell you that instead.
Work delivered in these environments, as an employee, contractor or through partners and vendors. Listed as engagements worked on, not as clients of this practice.
Alembic, IREN, Deutsche Telekom, Microsoft, Anthropic, NScale, CoreWeave, BoostRun and Novartis.
Atrium Health Wake Forest Baptist, Albert Einstein College of Medicine, Weill Cornell Medicine, Memorial Sloan Kettering and Mayo Clinic.
NOAA, NASA, Brookhaven National Laboratory and the Naval Postgraduate School.
AI factory and GPU cloud build-outs, semiconductor equipment, aerospace and aviation, automotive engineering, federal weather modeling, pharmaceutical research, academic medicine, clinical research computing, clinical pathology, defense postgraduate education, and a national laboratory.
Red Hat Certified Engineer across RHEL 3, 5 and 7. Sun Certified Systems Administrator. FreeBSD Certified Associate. Juniper JNCIS-SEC and JNCIS-ENT. Cisco CCNA. Microsoft MCSE and MCSA. NACI Public Trust clearance.
Master's degree in computer science. Two decades of practice since, most of it hands on, across every layer from the cabling to the application stack.
Separate from build and operate work: a scoped, evidence-backed assessment of what is actually exposed or underperforming, delivered as a findings register rather than a conversation.
Cluster, fabric and storage assessed against what the hardware should deliver. Benchmarked rather than asserted, with the configuration causes named and a remediation order.
Identity and access, data protection, logging and incident readiness, mapped to whichever framework applies to you. Every finding carries reproduction steps, an owner and a date.
A wider look at where money leaks, where work takes longer than it should, and where risk has never been priced. Run as a separate practice at Wirewalk AI.
The same engineers, under two names. The business-facing half of this practice — operating reviews, AI policy and governance, and AI infrastructure sold to people who do not use the word HPC — is at ai.wirewalk.com. One engineering practice; the two sites exist because the buyers read different words, not because there are two companies.
Practical guides to GPU and CPU performance, Linux and scientific software, AI training pipelines, parallel storage, InfiniBand and low-latency networking. Reproducible checks for AI factories and research centers, with security and recovery covered alongside the infrastructure.
AI pipelines · GPU and CPU benchmarks · Parallel storage · Low-latency networking · Linux software
Design Dell PowerScale and Isilon storage for AI and HPC pipelines: map client access, dataset layout, compute placement, capacity, and recovery requirements.
Choose and test Lustre stripe layouts for AI datasets, shared HPC files and checkpoints. Inspect actual file layouts and compare performance without global changes.
Diagnose slow PyTorch data loading on HPC clusters with per-rank timing, worker sweeps, storage checks, and reproducible GPU throughput measurements.
Measure distributed AI checkpoint stalls, storage contention, and restart time. Validate recoverable training state before tuning asynchronous saves.
Most of this does not need a discovery call first. If you already know what is wrong, or what you want checked, the engagement is listed with a fee and can be ordered directly — on a purchase order if that is what your finance team needs.
One problem, a day or a week, written up. A benchmark that disagrees with the datasheet, a mount that drops under load, a second opinion before a purchase order goes out.
InfiniBand, RoCE and Ethernet diagnostics with NCCL testing for GPU workloads, parallel filesystem setup and review, scheduler setup and workload review, or independent cluster acceptance testing.
Compliance readiness and full assessment against the framework that applies, and the remediation program that follows if you want the controls implemented rather than just listed.
Whether that is a cluster to design, a fabric that is not delivering what it should, a migration nobody wants to own, or a second opinion on someone else's architecture. If the honest answer is that you do not need outside help, we will say so.
Prefer to call? 917-217-7975. Details are used to reply to your enquiry. No mailing list, and nothing shared with vendors or partners.