Blog

Technical deep dives on agent security, DNS-based discovery, policy enforcement, and the architecture of trust.

ai-infrastructureinferencekv-cache

Under the Endpoint Picker, Part 1: One Request, Five Prices

The same prompt can wait half a second or several seconds for its first token, on the same model and the same GPU type. The difference is where the request was sent, and that is decided before any GPU sees it.

Igor RacicNovember 3, 2026Read
ai-infrastructuregpupcie

Under the Endpoint Picker, Part 2: What Sits Behind the Pod IP

A pod IP in an inference pool looks like one endpoint. Behind it are two computers joined by PCIe, talking through a ring buffer and a single register write. The same pattern runs NICs, NVMe drives and the data path in front of the cluster.

Igor RacicNovember 3, 2026Read
ai-infrastructuregpucuda

Under the Endpoint Picker, Part 3: Pay Once, Then Get Out of the Way

CUDA makes 685 calls into the kernel and GPU firmware to set up a device, then runs a dozen copies with none at all. That split between setup and the hot path is the best argument I know for how a data path, including the gateway in front of the GPUs, should be built.

Igor RacicNovember 3, 2026Read
ai-infrastructuregpuperformance

Under the Endpoint Picker, Part 4: Measure, Don't Assume

A hand-written GPU copy ran at two-thirds of CUDA's speed on the same hardware, with the same commands, over the same link. The cause was something nobody thought to check, and it is the same mistake round robin makes on every request.

Igor RacicNovember 3, 2026Read
ai-infrastructuregpuobservability

Under the Endpoint Picker, Part 5: Seeing Without Slowing Down

Watch the GPU doorbell from the wrong place and copy bandwidth drops from 55 GB/s to 0.02. Watch it from the right place and nothing changes. Inference observability follows the same rule.

Igor RacicNovember 3, 2026Read
ai-infrastructuregpudpu

Under the Endpoint Picker, Part 6: When the GPU Stops Asking the CPU

A single GPU thread can ring its own doorbell and start a data transfer with no CPU involved. It's a lab experiment today, but it shows where the data path is heading, and why policy has to live somewhere the data still passes through.

Igor RacicNovember 3, 2026Read
ai-infrastructurekubernetesinference

If Networking Is 90% of Agentic Infrastructure, the Network Should Be the Platform

How F5 BIG-IP Next for Kubernetes runs a GPU cloud or AI factory as one programmable platform, from picking the site down to picking the GPU.

Igor RacicNovember 2, 2026Read
biocomputingorganoid-intelligenceagent-security

GPUs Get Slower Under Load. Living Compute Changes Under Load.

This is Part 1 of a series reporting a lab study that placed retrying AI agents in front of a simulated GPU backend and a digital twin of living neural tissue, with a gateway that could set the pace of stimulation. The twin follows trends seen in recordings of real organoids, every figure here comes from that software model, and Part 2 will report the analysis of those recordings.

Igor RacicSeptember 19, 2026Read
agent-securitypolicyarchitecture

The Agentic Control Plane

A model can be fully safe and the system built around it still isn't, once it can call tools, invoke MCP servers, and touch infrastructure. The security boundary moved to the execution path, and it needs a control plane of its own.

Igor RacicSeptember 15, 2026Read
rdmanetworkinggpudirect

The Network Is Part of the Computer Now

Distributed AI training and inference live or die on the fabric between GPUs. A practical tour of RDMA, RoCEv2, GPUDirect, and NCCL, and why the network stopped being a connectivity layer.

Igor RacicAugust 30, 2026Read
mcpmcp-appsa2ui

The Agent UI Is Now Part of Your Attack Surface

MCP's stateless release candidate makes agent UIs a first-class part of the protocol. Most teams are about to treat this as a widget feature. It is a trust boundary.

Igor RacicJuly 17, 2026Read
dns-aidagent-securitycel

You Can't Prompt Your Way to Agent Security

LLMs repeat the same architectural mistake that gave us buffer overflows. Agent security requires deterministic policy enforcement outside the model's reasoning loop.

Igor RacicMarch 23, 2026Read
naasnetworkingcloud

The Double-Edged Sword of NaaS: A Deeper Look

Network-as-a-Service promises agility, lower CAPEX, and faster time to service. But the hidden trade-offs — lock-in, loss of control, and unpredictable long-run costs — deserve a closer look.

Igor RacicAugust 9, 2023Read
pythonnetworkingbandwidth

Analyzing Bandwidth Consumption Using Python Classes: A Deep Dive

A class-based Python solution for analyzing per-application bandwidth consumption across enterprise sites. Practical OO design applied to a real network operations problem.

Igor RacicAugust 8, 2023Read
dpireverse-proxynetworking

Deep Packet Inspection (DPI) vs. Reverse Proxy: Navigating the Nuances of Network Management

DPI and reverse proxies both shape modern network management — but they solve different problems. A side-by-side look at granular control, intrusion mitigation, load balancing, and SSL termination.

Igor RacicAugust 7, 2023Read
networkinglayer-3cloud

Beyond Basics: Why Layer 3 Networks Fall Short in Modern Cloud Landscapes

Layer 3 routing was built for a world of IP-addressable endpoints. In a cloud landscape of services, identities, and policies, IP-centric design is no longer enough.

Igor RacicAugust 7, 2023Read
multi-cloudmcnnetworking

Multi-Cloud Networking — Why Has It Taken So Long?

Private, public, multi-cloud, hybrid. The terms blur, but the networking problem doesn't. Why traditional networking still struggles with cloud-native infrastructure — and what an MCN platform must do.

Igor RacicJanuary 27, 2023Read