Mohammad Ayan
Lead MLOps At Rakuten Symphony
Bengaluru, India
Actions
I'm an AI infrastructure engineer who builds GPU platforms that teams actually ship on. I've designed and scaled multi-tenant, hardware-isolated AI infrastructure for orgs like Rakuten, PhysicsWallah, GMR Group, covering everything from distributed training & inference to disaggregated serving and speculative decoding on Kubernetes.
I also hold top industry certifications including NVIDIA Certified Professional – AI Operations (NCP-AIO) and Certified Kubernetes Administrator (CKA).
I'm a open-source contributor and built Fake Network Operator from scratch, a full Kubernetes operator to simulate InfiniBand and RDMA networking, because the tooling to test GPU networking without six-figure hardware simply didn't exist.
Based in Bengaluru, Karnataka, India.
Building in the open.
Links
Badges
Area of Expertise
Topics
$0 InfiniBand: How I Simulated a $200K NVIDIA DGX Network Stack on a Laptop with KinD
Every AI platform team has the same dirty secret: they're burning $400/hr on cloud GPUs, not to train models, but to test whether a Helm chart correctly requests RDMA devices, scheduling configs actually work, and NetworkAttachmentDefinitions don't break in production.
Networking layer of Distributed AI: InfiniBand, RDMA, SR-IOV, ConnectX NICs did not have any local simulation tools. Fake GPU Operator dealt with GPUs. KWOK dealt with nodes. Networking? Either you were DGX or blind-shipping to test.
I built Fake Network Operator to kill that entire cost line.
What changes for you after this talk:
• This is where you realize how the GPU sharing systems (Volcano, Kueue) really fit into the networking infrastructure – RDMA, Infiniband, SmartNICs, hardware isolated Virtual Functions using SR-IOV without a trace of any of the hardware.
• You schedule real pods that request real extended resources, get assigned real device IDs, attach secondary network interfaces, and run infiniband (IB) diagnostic commands, all on your laptop. Your workloads don't know the difference.
• Your platform team ships infrastructure changes with the same fast feedback loop they expect from application code. No more 3-week hardware procurement blocking a one-line YAML fix.
In 90 seconds, I will demo a full InfiniBand + RDMA + SR-IOV cluster using KinD, explain how the three layers fool Kubernetes into thinking it is running on bare metal, and provide you with a permanent helm installation of all of the above.
Have you ever spent three weeks waiting for InfiniBand access to test a one-line YAML patch? If so, this talk is for you!
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top