Work

Systems we've designed and built.

Each project comes with a design document and tests. Where the trade-off matters, the same policy is built in more than one dataplane, so the choice is measured rather than guessed.

01 Fabric

Fabric and transport

GPU fabric

RoCEv2 fabric generator

Designs a rail-optimized, non-blocking two-tier RoCEv2 fabric and turns one JSON spec into cabling, FRR, QoS intent and per-host NCCL setup. Refuses specs that break the design rules. Python, C and Go ports produce byte-identical output.

  • RoCEv2
  • FRR / BGP
  • NCCL
  • Python / C / Go

RDMA security

RoCEv2 firewall and telemetry

Parses the InfiniBand transport header and checks transport type, partition key, peer and queue pair. Counts ECN marks and CNPs per sender and queue pair. Built as a kernel module, in XDP and in DPDK.

  • RoCEv2
  • eBPF / XDP
  • DPDK
  • Netfilter

Collective transport

Transport design reviews

Design documents for an OCCL-NET collective network plugin and the OCP Multipath Reliable Connection (MRC) API, covering how the transport maps onto RDMA hardware.

  • NCCL
  • RDMA verbs
  • OCP MRC

02 Datapath

Offload and accelerate

DPU offload

BlueField flow agent

A proactive OpenFlow 1.3/1.4 and gRPC agent for the BlueField eSwitch. It compiles wildcard rules into mlx5 hardware-steering template tables with async commits, rollback and a journal, and gates traffic until every rule is in hardware.

  • DPDK
  • rte_flow HWS
  • OpenFlow
  • gRPC

Hardware offload

Vendor-agnostic OVS-DPDK offload

An offload architecture in which one control path (OpenFlow, OVS, rte_flow) runs on BlueField, Pensando and Intel hardware. Flows the hardware cannot take stay in software instead of being dropped.

  • OVS-DPDK
  • rte_flow
  • SDN

NFV on DPU

NFV pod on BlueField-3 virtio-net

Places a session border controller’s DPDK media pipeline on DPU-emulated virtio VFs. Classification, NAT hits, IPsec and DoS move to the eSwitch and Arm cores, and media stays in the pod.

  • DOCA
  • OVS-DOCA
  • virtio-net
  • Kubernetes

03 Workloads

Secure and tune

AI workload security

aiguard

eBPF container security for AI workloads. 22 BPF LSM programs and 4 cgroup socket programs enforce exec, read-only model, secret and egress allowlists, block escapes, and learn policy from audit mode.

  • eBPF
  • BPF LSM
  • libbpf / CO-RE

Network functions

Firewall, NAT and rule engine in three dataplanes

A stateful firewall, an IPv4 NAT gateway, a SYN filter and a runtime rule manager, each written as a netfilter kernel module, in XDP/TC and in DPDK. Rule updates are atomic, so no packet sees a half-updated rule set.

  • Netfilter
  • XDP / TC
  • DPDK
  • C

Edge protection

XDP firewall

An XDP ingress firewall with IPv4/IPv6 LPM blocklists, every XDP verdict as a rule action, a per-source SYN-flood guard, and rule reload without detaching.

  • XDP
  • eBPF
  • libbpf

Also built

Services and applied AI

Infrastructure services

dnsd

An authoritative and caching-forwarder DNS server in C11 with only libc. It serves UDP and TCP with EDNS(0), and is tested with unit, fuzz and end-to-end tests under sanitizers.

  • C11
  • DNS
  • Fuzzing

Rust and DPDK

rte-flow-rs

Safe Rust bindings over DPDK rte_flow that install a hardware rule and read its counters, with RAII teardown of flows.

  • Rust
  • DPDK
  • bindgen

Applied AI

AI phone receptionist

Answers calls for a home-service business around the clock. It books visits, transfers emergencies to the on-call technician and texts the owner a summary of each call, using speech-to-speech AI.

  • Node.js
  • Twilio
  • Realtime voice AI

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