StrataPath.ai

Services

Start at the layer you need most.

Some teams need the whole stack designed. Others need one layer fixed. Each service stands alone, and all three fit together.

01 Fabric

Fabric and transport

Get collective traffic off the critical path. We design the GPU network and the transport above it, and generate the configs so what is built matches what was designed.

What you get

  • Rail-optimized, non-blocking RoCEv2 leaf-spine design with ECN, DCQCN and PFC
  • Generated cabling plans, switch configs, host setup and NCCL settings, with cabling checks
  • Transport reviews: NCCL net plugins, multipath RDMA (OCP MRC), RoCE firewalling and per-QP telemetry

Best when

  • You're planning or expanding a GPU cluster
  • Collectives are slower than the hardware should allow
02 Datapath

Offload and accelerate

Move packet work off host cores. We build kernel-bypass fast paths and push classification, NAT, IPsec and switching into NIC and DPU hardware.

What you get

  • DPDK pipelines with run-to-completion workers, RSS and bounded exception rings
  • BlueField DPU and ConnectX offload: rte_flow hardware steering, OVS-DOCA, conntrack and inline IPsec
  • SR-IOV, virtio-net emulation and Kubernetes integration for NFV pods
  • Vendor-agnostic designs that fall back to software when hardware cannot take a flow

Best when

  • Network functions are running out of host cores
  • You are adopting DPUs and need a plan for what moves where
03 Workloads

Secure and tune

Protect and speed up what runs on top. We enforce policy in the kernel with eBPF, and tune the platform until the numbers hold under load.

What you get

  • eBPF / XDP firewalls, SYN-flood guards, stateful conntrack and NAT
  • BPF LSM container hardening for model files, secrets and egress
  • Profiling of PCIe, NUMA, queues and cores, with benchmarks and recommendations

Best when

  • AI workloads handle sensitive models or data
  • Throughput or tail latency misses its target

Ways to work together

Choose the depth that fits.

Performance audit

A short, focused engagement that profiles one system and ends with findings, a design sketch and a recommended next move.

[Duration] · [Price]

Design and build

Hands-on work across one or more layers, from design document to tested, benchmarked code your team owns.

[Duration] · [Price]

Advisory retainer

Ongoing senior review for infrastructure and platform teams as hardware, kernels and workloads change.

[Duration] · [Price]

Questions

Before you get in touch.

Do you write the code, or only the design?

Both. Each engagement can end with a design document, working code in C, C++, Go, Rust or Python, tests, and benchmarks against a baseline.

Which hardware do you work with?

Most of our work is on NVIDIA BlueField DPUs and ConnectX NICs with DPDK, DOCA and OVS. Designs use standard interfaces such as rte_flow, so they can carry over to other NICs and DPUs.

Should this run in the kernel, in XDP or in DPDK?

It depends on what the system has to do. We often build the same policy in more than one dataplane and measure, so the choice rests on numbers.

How do you handle confidential data?

[Describe your confidentiality terms, security practices and any certifications you hold.]

Find the bottleneck before you buy more hardware.

Bring a workload and a target. In one call, we'll help you work out where the time goes and which layer to fix first.

Book a technical call