CodiLime

We don't show up to these things just to get our logo on a leaflet. When we join a conversation, it's because we know we can move it somewhere new. CodiLime X ONUG. Fall 2026. More to come…

We taught the Kubernetes scheduler what workloads actually do. Same three Jobs finished in 16 minutes instead of 37. The pipeline: Prometheus collects real utilization → an ML model learns each workload's pattern and forecasts it → a node evaluator scores every node on those forecasts → a custom scheduler plugin (hooked into Kubernetes' own extension points) asks, per node, "can this Pod run here, and how well?" No fork of the scheduler. No patched core. Just a plugin the framework already allows, scoring on learned behavior instead of static declarations. There's no AI chatbot in this story, and that's the point. The differentiation isn't a model; it's the ML engineering. Full breakdown here: https://lnkd.in/dkDczz5V

🚀 Do you enjoy working with complex networks, infrastructure automation, and modern AI hardware? We’re looking for engineers who want to build, test, and automate large-scale infrastructure, solve challenging networking problems, and validate hardware powering the next generation of AI. Depending on the role, you’ll work with technologies and concepts such as SASE, Go (Golang), Python, C, network automation, Terraform, Ansible, EVPN/VXLAN, BGP, data center networking, hardware & network validation, and CI/CD. We’re currently hiring for: 🔐 Mid/Senior SASE Automation Engineer https://lnkd.in/d7ZrvNYM 🤖 Golang Software Engineer with experience in Python/C https://lnkd.in/dKYbHkqV 🌐 Senior Network Deployment Engineer https://lnkd.in/dm9uePG7 🔐 Network & Hardware AI Validation Engineer https://lnkd.in/dyM-bvQT If you’re looking for a role where you can work on technically challenging projects, build production-grade solutions, and collaborate with experienced engineers - take a look at our openings. Your next challenge is waiting for you at CodiLime. 🚀

Two apps. Two peak-sized reservations. Two nodes. Except their peaks never happen at the same time. One app peaks during business hours. The other peaks in the evening. Scheduled by their declared requests, each demands its own node, because the scheduler only sees the static number, not the shape. Teach it the shape, and it sees what you can see in this chart: the two curves fit under one node's ceiling, because they never spike together. Two nodes' worth of declared demand, served by one. That's the scheduler scoring on learned reality instead of a manifest. How we built it, find out here: https://lnkd.in/dkDczz5V

CodiLime is at Black Hat USA this week in Las Vegas. Our Chief Revenue Officer, Arindam Guha, is there with one shift on his mind: security is moving down into the infrastructure layer itself. The network, the data plane, the hardware, not only the software running on top. That is the ground we know. We build and secure the layer where networking meets software: vendor-neutral data center modernization, automation, and AI-ready infrastructure that is secure by design. If you are at Black Hat too, Arindam would love to compare notes over coffee. Send him a message. #BHUSA #Cybersecurity #Infrastructure

Your Kubernetes cluster is half-empty and still running out of memory. Both things are true at once, and here's why. You declare a memory request for each workload, and the scheduler reserves that much 24/7. So you set it near the peak to stay safe. But your app only hits that peak for a few hours a day. The rest of the time, the scheduler is guarding capacity nobody's using (the shaded gap in this chart). Set requests lower to reclaim it, and you risk everything spiking at once. The number is a guess. The cluster pays for the error. We built a PoC that replaces the guess with a learned forecast. Read about it here: https://lnkd.in/dkDczz5V

As Shakespeare once didn’t say: to build, or not to build, that is the network automation question… Every network automation team eventually hits this same question, and most treat it as one big decision to win, when it's really a series of smaller ones to get right. Our new issue of Up Next argues build vs. buy was never really a purchasing question. It's a question about where your engineering effort is best spent. Write it yourself, and you get an exact fit, plus every hour of maintaining it. Buy a platform, and you get speed and a support line, plus someone else's assumptions about your network. Inside this edition: → The complete guide to making the right call → Our build and buy advocates going head-to-head → Two real success stories from both sides of the debate → Why agentic AI raises the price of a weak foundation Read it and subscribe below. 👇 #NetworkAutomation #BuildVsBuy #CodiLime

The CPU does exactly three things in our AI firewall. Then it walks away and never comes back. Phase 1, runs once: allocate the shared buffers, capture the GPU's CUDA graph, launch the firewall app on the DPU. Setup done. Phase 2, runs forever: the DPU and GPU handle every packet between themselves. No CPU dispatch loop. The GPU replays its own captured graph and schedules its own work. Most "AI in infrastructure" means wiring a model in at the application layer, where everyone has the same tools. Designing a data path the CPU can leave entirely is a different altitude of engineering, and far fewer teams can build there. That's the layer this one lives on. Want to know how to get there? Read this: https://lnkd.in/dPCeZh2s

Here's the entire life of a packet in our AI firewall. Read the six steps and notice what's missing. Arrives at the DPU. Relevant data extracted. Written to GPU memory. Model runs. Verdict read back. Packet dropped or forwarded. No host CPU. Anywhere. It serves the web application (the thing it's good at), while the firewall runs as a private conversation between two peripherals. Bonus: hostile traffic gets inspected and discarded before it ever touches host memory. The system never handles the data it's filtering. Full breakdown here: https://lnkd.in/dPCeZh2s

For the past month, we traded some screen time for sports shoes. Using a dedicated app, our team logged their favorite physical activities, competed for kilometers, steps, and most importantly - raised money for charity! 🤝❤️ This is #TeamUpToWin and #ActToDeliver in its purest form. Check out what we achieved together: 👥 42 participants gave it their all 🏋️♂️ 1927 sports activities logged 🛣️ 9790 km covered across all activities 👟 6 118 887 steps taken! Huge applause to everyone who broke a sweat to support a great cause. You rock!