About
I design and operate large-scale cloud infrastructure and platform systems — with 12+…
Services
Articles by Hitesh
Activity
2K followers
-
Hitesh Sahu shared thisI’m heading to NVIDIA GTC Berlin 2026 and, as usual, I ended up building something because the existing solution wasn’t quite how I wanted to use it. Behold an improved GTC Agenda: 👉 https://lnkd.in/ddmAwSWv The problem is that GTC has a lot going on. NVIDIA’s official agenda covers sessions across CUDA, AI infrastructure, agentic AI, inference, robotics, physical AI, open models, and more. The conference itself runs October 20–22, with the keynote on October 21. It's hard to find: “What is happening at 2 PM, where is it, and what should I attend next?” The official session catalog is useful for static browsing, but it wasn't great for quickly building a complete personal schedule. Sessions are presented as cards, and you have to search, filter, open sessions, and add them to your calendar individually. Also, data is paginated across multiple pages with a load more button. So I built my own tool by reverse engineering NVIDIA's own backend API. From there, it became a small engineering project: NVIDIA session data → REST API → agenda → filtering/search → calendar export → Google Calendar What I changed Instead of treating the agenda as a collection of pages, I turned it into a single searchable view of the event. Single-page unified searchable agenda page with rooms route thanks to Google Maps Integration. • 📅 Filter by day and session type • 🕐 Sort the entire agenda by time • 📍 Quickly see where each session is happening The live session is highlighted, and past sessions will be greyed out. (auto-refreshes every 60 seconds) • 🟢 See what's live, upcoming, or already finished • 📱 Designed for quickly checking the agenda between sessions • 📆 Export the agenda to Google Calendar • 🎨 Keep session types visually organized in the calendar And because the agenda can change, the page fetches the current data rather than being a static list. Also, data is cached in your browser, so even if there is no internet, you will still see the last fetched agenda list. if you like to know the technical insight, I have created a post here: https://lnkd.in/ddEqWxhc Maybe I'll see some of you in Berlin! 👋 #NVIDIAGTC #GTCBerlin #GTC2026 #NVIDIA #AI #ArtificialIntelligence #AIInfrastructure #CUDA #AgenticAI #GPU #MachineLearning #SoftwareEngineeringGTC 2026 Agenda – Full GTC Berlin Session ScheduleGTC 2026 Agenda – Full GTC Berlin Session Schedule
-
Hitesh Sahu shared thisI’m Going to GTC Berlin. So I Built My Own Agenda ToolI’m Going to GTC Berlin. So I Built My Own Agenda Tool
-
Hitesh Sahu shared thisHappy to share that I’ve cleared my German B1 exam! 🇩🇪 This was something that had been sitting on my to-do list for a long time. And honestly, I had been avoiding it for just as long. I’ve realised that I work better when I have a bit of pressure. So instead of waiting until I felt completely ready, I decided to face my fear and book the exam, even though I wasn’t fully prepared. That decision gave me something I had been missing: a deadline. Once the exam was booked, postponing was no longer an option. I had to show up, prepare, and give it my best. The biggest takeaway for me wasn’t just passing the B1 exam. It was learning that sometimes you don’t need to feel ready before you start. You just need to commit. One more item off the list. ✔️ And one step closer to feeling at home in Germany. 🇩🇪 I’ve also put together and shared the notes I used while preparing, in case they’re useful to anyone else on the same journey: https://lnkd.in/eMs-8Ppd And if you’re somewhere on the same German-learning journey and have been putting off that exam, I hope this gives you a little push. You might not feel ready. Book it anyway. Give yourself a deadline, and start. 💪
-
Hitesh Sahu shared thisNVIDIA is no longer just the company behind CUDA and GPUs. Over the past few years, it has become one of the largest contributors to the open AI ecosystem. Across GitHub and Hugging Face, NVIDIA now publishes projects spanning foundation models, inference, Kubernetes, #robotics, HPC, data science, and developer tooling. But "open" doesn't always mean the same thing. The repository count sounds great, but the license is the contract. If you're evaluating technologies or building products, understanding the license matters as much as understanding the technology. Here's the map, grouped by licensing model rather than technical domain. 🟢 Open Source Permissive open-source licenses with community-visible source code. • Kubernetes & Cloud Native – GPU Operator, Network Operator, Device Plugin, MIG Manager, GPU Feature Discovery • HPC & Networking – NCCL, NVSHMEM, open tooling around GPUDirect-enabled communication • CUDA Libraries – CUTLASS, CUB, Thrust, cuCollections • Data Science – RAPIDS (cuDF, cuML, cuGraph, cuSpatial) • Agent Frameworks – NemoClaw • Upstream Contributions – Kubernetes, Linux kernel, PyTorch, pandas, scikit-learn, Docker, OpenUSD 🟣 Open Weights: Model weights, inference examples, and evaluation assets are published. Licenses vary by model family. • Foundation Models – Nemotron, embeddings, rerankers, VLMs, ASR/TTS. Weights and evaluation assets are published, but licenses vary by model family. 🟡 Mixed: A combination of open-source and proprietary components. • Inference – Triton, TensorRT-LLM, Dynamo, vLLM, SGLang (TensorRT itself remains proprietary) • Robotics – Isaac Lab & Isaac ROS (open), Isaac Sim (licensed) • Computer Vision – CV-CUDA (open), DeepStream (binary SDK) • Omniverse – OpenUSD is open, Omniverse Kit is licensed 🔴 Proprietary: Free to use or commercially licensed, but not open source. • Developer Tools – Nsight profilers, CUDA debugging, performance analysis tools For platform and infrastructure engineers, the biggest opportunities aren't in training models. They're in building the systems around them: GPU scheduling, inference infrastructure, distributed networking, and high-performance communication. What's the NVIDIA project you've actually used in production? #OpenSource #NVIDIA #GPUInfrastructure #Kubernetes #AIInfrastructure #MLOps
-
Hitesh Sahu shared thisImagine there's a big cooking contest for AI robots. Lots of companies build robot chefs. There's one from America called Claude, one called GPT, and a bunch of others. For a long time, the American robot chefs kept winning the contests. Now there's a new robot chef from China called Kimi K3. Last time, Kimi's older brother came in 18th place — not great. But this new Kimi just entered the contest and won 1st place! It jumped from 18th to 1st in one go. That almost never happens. How does the contest work? People taste food from two robots without knowing which robot made it, and pick the one they like better. Kimi's food got picked 76 times out of 100. Claude Fable 5's got picked 63 times out of 100. So Kimi really did win fair and square at this contest. ⚠️ This is one leaderboard (Arena's Frontend Code). On overall performance and benchmarks like DeepSWE, the closed frontier models still lead. ⚠️ The open weights everyone's excited about are promised for July 27, not released yet. Announcing you'll share the recipe isn't sharing the recipe. But 18th → 1st in a single generation is the trend line worth watching. The gap between open and closed models used to be measured in years. Now it's months. And if a 2.8-trillion-parameter MoE with a 1M-token context window really does go open next week, self-hosting frontier-class models stops being a thought experiment — and becomes a GPU infrastructure problem. My favorite kind. 🏗️ Would you route frontend generation to a model you can self-host, or does API convenience still win? #AI #LLM #OpenSource #GPUInfrastructure #KimiK3 #AIEngineering
-
Hitesh Sahu shared thisAmazon Web Services (AWS) just launched SimuLearn learning plan badges: verifiable Credly badges you earn by solving AI-powered customer simulations and building real solutions in a live AWS environment. Unlike traditional courses, each simulation makes you: 1️⃣ Gather requirements from a virtual customer 2️⃣ Design the AWS architecture 3️⃣ Build it in a real AWS environment 4️⃣ Pass automated validation Complete every assignment in a learning plan → AWS auto-issues a Credly badge for LinkedIn or your resume. I did a quick cost/benefit breakdown before jumping in: 💰 Cost • FREE (just an AWS Builder ID): - AI Practitioner - Cloud Practitioner • Skill Builder subscription ($29/month or $449/year): - AI Architect - Solutions Architect - Security - Networking - Serverless Developer, - Machine Learning - Data Analytics - Industry tracks (Financial Services, Healthcare, Manufacturing & Automotive) ⏱️ Time • Each simulation runs roughly 1 hour, end to end • A full learning plan is typically 10–20 hours — one focused month on the $29 plan covers a badge comfortably ✅ Advantages • Hands-on validation, not multiple choice — you build in a real AWS console • Auto-issued via Credly, verifiable by recruiters • Sandboxed environments — zero risk of surprise AWS bills • The paid sub also unlocks official practice exams and Builder Labs, so it doubles as cert prep ⚠️ Disadvantages • Badges ≠ certifications — they won't renew or replace an AWS cert • Hiring managers still weight them below actual certs (for now) • Scenarios are guided; real production troubleshooting is messier • $449/year is steep if a badge is all you want 👥 Who benefits most • Career switchers who need demonstrable hands-on proof before their first cloud role • Students / early-career engineers building a credential trail on a budget (start with the 2 free ones) • Cert holders wanting a hands-on refresher between recertification cycles • Domain specialists — the Financial Services, Healthcare, and Manufacturing & Automotive tracks are rare industry-specific credentials My take: start with the free AI Practitioner track, and only pay if you're pairing it with cert prep anyway. Anyone tried SimuLearn yet? Curious how the AI customer conversations hold up in practice. #AWS #CloudComputing #SkillBuilder #CloudSkills #ContinuousLearning
-
Hitesh Sahu shared thisFor fifteen years, we shipped libraries you import. Now we're shipping capabilities that an agent can absorb: skills, MCP servers, DESIGN.md files, and memory systems. The skill pack is becoming the new npm package. And notice where the gravity sits. This week's top 10 trending GitHub repos solve the same unglamorous problem, giving agents durable memory and reusable knowledge. Persistent context is the land grab right now, and nobody has clearly won it. The pattern is hard to miss. 🦾 SKILL PACKS & AGENT CAPABILITIES Reusable expertise and new abilities you hand straight to an agent. ➡️ skills: https://lnkd.in/eF32pHVp - Skills for real engineers, straight out of Matt Pocock's .claude directory. ➡️ Anthropic-Cybersecurity-Skills: https://lnkd.in/eqZeY3gU - 817 structured cybersecurity skills for AI agents (a community collection, not an official Anthropic release). - Mapped to MITRE ATT&CK, NIST CSF 2.0, MITRE ATLAS, D3FEND, NIST AI RMF, and MITRE F3. Works with Claude Code, Microsoft Copilot, Codex CLI, Cursor, Gemini CLI, and 20+ more. ➡️ Agent-Reach: https://lnkd.in/eV2iBZrZ - Gives AI agents "eyes" for the internet. Search & read across Twitter, Reddit, Inc., YouTube, GitHub, bilibili Group, and Xiaohongshu. One CLI, no API fees. 🧠 PERSISTENT CONTEXT & MEMORY Durable, structured knowledge agents can carry between sessions. ➡️ codebase-memory-mcp: https://lnkd.in/eRrQHY_p - High-performance MCP server for code intelligence. Indexes codebases into a persistent knowledge graph: average repos in milliseconds, 158 languages, sub-ms queries, 99% fewer tokens. Single static binary, zero dependencies. ➡️ design.md: https://lnkd.in/eQ78f7B3 - A format for describing visual identity and design systems to coding agents. A DESIGN.md gives agents a persistent, structured sense of how a product should look and feel. 🎬 AI CREATION STUDIOS AI is moving past code into media generation. ➡️ OpenMontage: https://champ.ly/FDSY0kR6 - The first open-source, agentic video production system. 12 pipelines, 52 tools, 500+ agent skills. Turns your AI coding assistant into a full video production studio. ➡️ voicebox: https://lnkd.in/efQQY86A - Open-source AI voice studio. Clone, dictate, create. 📊 AI-POWERED APPS & TEMPLATES Verticals where agents do the heavy lifting. ➡️ daily_stock_analysis: https://lnkd.in/euy9Nam3 - LLM-powered multi-market stock analysis. Multi-source market data, real-time news, a decision dashboard, automatic notifications, and free scheduled runs. ➡️ ai-website-cloner-template : https://lnkd.in/ewWvBKUJ Clone websites with a single command via AI coding agents. ➡️ penpot : https://lnkd.in/e59vSRHT - Open-source design tool for design and code collaboration. So the question in 2026 isn't "what app do I build" anymore. It's "What capability do I hand the agent?" And that's exactly where the race is still wide open. #GenAI #GitHub #AIAdoption #DeveloperTools
-
Hitesh Sahu shared thisI created a tool for spinning up a Slurm Workload Manager cluster with an option to use a real CUDA GPU and run an #HPC workload on it: ⭐ GitHub: https://lnkd.in/ewtSnTru If you need to run batch jobs for HPC, then you must be aware of SLURM. It is like Kubernetes for HPC clusters, but focused on batch jobs and resource scheduling. One thing I noticed was that many common Slurm operations require remembering a long list of commands: • sinfo • squeue • sbatch • scontrol • sacct (when available) • ...and many more. For someone new to HPC or AI infrastructure, the learning curve can be surprisingly steep. Going through the quick start guide does not make it better, and it's a pain to start a cluster. So I started building Caravan 🐫. Caravan is an open-source CLI built with Golang lang that provides a simpler interface for interacting with NVIDIA Slurm clusters while staying true to the underlying Slurm ecosystem. Instead of memorizing dozens of commands, you get a consistent developer-friendly experience. You can install quickly on a MacBook: $ go install https://lnkd.in/eHV3FH2v Now you can start a Slurm cluster, check the status, and submit jobs & tear down easily: - $ caravan cluster up - $ caravan cluster status - $ caravan submit workloads/submit_example.sh - $ caravan cluster down By default, it uses a fake GPU so that you can run it on MacOS without a GPU. But you can assign 1 GPU in the C1 compute node to run workloads. The same setup as we usually have: a workstation with 1 GPU installed # GPU Load: need real GPU + Linux/WSL2 $ CARAVAN_GPU=real ./caravan submit workloads/gpu_example.sh The long-term vision is to make working with GPU clusters as intuitive as using modern cloud CLIs, while still exposing the full power of Slurm when you need it. This is the first release & this project is also part of a growing ecosystem I'm building around AI infrastructure: 🦝🐾 squint → GPU-aware Slurm monitoring TUI - https://lnkd.in/eYAxYHJ2 🔎 gpu-lens → GPU diagnostics and inspection - https://lnkd.in/ekextmUv 🐫 Caravan → A friendly CLI for Slurm clusters - https://lnkd.in/ewtSnTru If you work with AI training, HPC, or GPU infrastructure, I'd love to hear your feedback. Open to feature requests, ideas, and contributions! #OpenSource #GoLang #HPC #Slurm #GPU #AIInfrastructure #MLOps #DevOps #CloudNative #Linux #CLI
-
Hitesh Sahu shared thisWhen I was younger, I'd constantly check the progress of a game download, hoping there were enough seeds for it to finish. Now I watch AI agents with the same anticipation, hoping they'll complete the task without throwing any errors or running out of tokens. Different time, different tech but both feeding on gpu & giving same level of anxiety.
-
Hitesh Sahu liked thisHitesh Sahu liked thisAre you confused about which Geely is which? Here’s a post for you. First, stop saying ”Geely”. Geely Holding Group is not 1 company or 1 brand. The car brand Geely is from the company named Geely Auto. Which is a subsidiary of Geely Auto Group that also includes Zeekr and Lynk & Co. And Geely Auto Group’s main shareholder is the privately owned Geely Holding, which also is a majority shareholder in OEMs like Volvo Cars (NOT Volvo Group…), Polestar, Lotus etc. Its really not that complicated, folks. Just like the company that sell cars branded Ford is named ”Ford Motor Company” and Volkswagen Group has a car brand named Volkswagen aside other brands. Hence, there’s a few different ”Geely”-named Linkedinpages. Such as Geely Auto Europe that then naturally sells Geely branded cars. While Geely Holding EuropeGeely Holding Europe deals with our investments in various areas. Got it? #geely #branding
-
Hitesh Sahu liked thisHitesh Sahu liked this🚀 A bit late with the update, but officially a Kubestronaut - and the 129th in Poland! 🇵🇱 I actually finished all five Kubernetes certifications back in April, but somehow never got around to sharing it here 😅 ☸️ Kubernetes and Cloud Native Associate (KCNA) 🔐 Kubernetes and Cloud Native Security Associate (KCSA) ⚙️ Certified Kubernetes Administrator (CKA) 🧑💻 Certified Kubernetes Application Developer (CKAD) 🛡️ Certified Kubernetes Security Specialist (CKS) Five certifications covering cloud native fundamentals, administration, application development, and security. A big thank you to KodeKloud and Mumshad Mannambeth for the courses that supported my preparation. The dedicated Kubestronaut Learning Path helped bring structure to the journey across all five certifications. For anyone planning their own journey, here are a few additional resources worth exploring: 🔹 Killer.sh - exam simulators for practising CKA, CKAD, and CKS scenarios 🔹 Kubernetes: Up and Running - a practical foundation for understanding and working with Kubernetes. 🔹 Kubernetes in Action - a deeper look at deploying and managing applications, with coverage of Kubernetes architecture and security. Happy to join the Kubestronaut community and keep building on what I’ve learned. #Kubestronaut #Kubernetes #CloudNative #CNCF #KodeKloud #DevOps #PlatformEngineering #KubernetesCertification #Certifications
-
Hitesh Sahu liked thisHitesh Sahu liked thisIf you’re an AI Infrastructure Engineer, here are 5 projects you must build. Trust me, these are some gold projects that will help you understand how AI systems actually run in production. 1. Dockerfile Generation using Local LLMs (Beginner) Use Ollama with an open-weight LLM to automatically generate Dockerfiles for different projects ↳ Tools: Ollama, Python, Hugging Face, Docker ↳ YouTube: https://lnkd.in/d36ZCJ4m 2. Build a Production LLM Serving Stack on Kubernetes Deploy an LLM with vLLM on Kubernetes and get hands-on with multiple replicas, routing, observability, model-aware routing and KV-cache management. ↳ Tools: vLLM, Kubernetes, Helm ↳ GitHub: https://lnkd.in/g_5EZnWV 3. Build an End-to-End DevOps + AIOps Project See how modern platforms monitor themselves, detect issues and automate operational responses at scale. ↳ Tools: Kubernetes, Monitoring, Observability, AIOps ↳ Project: https://lnkd.in/gY-yevBb 4. Build a GPU Autoscaling Inference System Deploy GPU-backed inference and generate concurrent traffic to see how your system automatically scales based on workload. ↳ Tools: KServe, Kubernetes, GPUs ↳ GitHub: https://lnkd.in/eGHpJKy9 5. Build Scale-to-Zero GPU Inference with KEDA This is a full end-to-end project, not just a KEDA example You’ll get hands-on with queue-based scaling, scale-to-zero, cold starts, GPU nodes and inference workloads. ↳ Tools: KEDA, Redis, Kubernetes, vLLM, GKE ↳ GitHub: https://lnkd.in/eg4YqUQ3 AI Infrastructure Engineering has a whole set of components and use cases you can master ~ from containers and Kubernetes to inference, GPUs, autoscaling, observability and AIOps. These projects are a great starting point. → If you have access to GPU resources, great. Build them, break them and benchmark them. → If you don’t, there are still ways to simulate the environment and reproduce parts of these projects on your laptop. And if even that’s not possible, just read through the documentation and architecture.. You’ll still learn a lot of concepts that will come in handy when you eventually work with these systems. Save this for later and give one of these projects a try this weekend!!
Experience & Education
Licenses & Certifications
Join now to see all certifications
Volunteer Experience
-
Open Source Developer
GitHub
- Present 13 years
Open source contribution with over 400 followers, 100+ repos: https://github.com/hiteshsahu
-
Administrator
Stack Overflow
- Present 12 years 11 months
Answered 600+ technical questions helping 40 Million+ Developers worldwide. With 40,000+ reputation points & 1200 rank worldwide, I am helping other developers as Stackoverflow Admin: https://stackoverflow.com/users/2252113/hitesh-sahu
Languages
-
English
Professional working proficiency
-
Hindi
Native or bilingual proficiency
-
German
Elementary proficiency
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content