मुख्य कंटेंट तक स्किप करें

अभी (October 2026)

Inspired by Derek Sivers' /now movement

🎯 Current Focus​

Building: एक GPU inference platform जो $0 idle पर scale करती है — Karpenter + KEDA HTTP add-on + metrics-api scalers across 5 model types (LLM, embedding, OCR, vision, multimodal)। Running on Blackwell (B200/B300), L40S, और L4 spot fleets.

Optimizing: Post-credits cloud cost strategy — bills को baseline (commit) vs variable (on-demand) में split करना। 1-year no-upfront Compute Savings Plan + targeted RIs databases के लिए। ODCR business-hours GPU capacity के लिए।

Learning: NVFP4 quantization on Blackwell, FlashInfer attention backends, vLLM 0.6+ continuous batching improvements, और Karpenter 1.0 GA release।

Writing: "Running GPUs in 2026" doc series (6 providers) खत्म करना, EKS Security Hardening guide (5 layers), और AWS Account Setup playbook (5 guides)।

🔧 Active Experiments​

  • SOCI parallel-pull + FastImagePull sub-50s GPU cold starts के लिए
  • Value-based GPU taints (nvidia.com/gpu=workload-x) perfect NodePool isolation के लिए
  • vmagent + vmauth edge device metrics के लिए (20 GB disk queue > Alloy's 8-hour WAL)
  • Business-hours ODCR via EventBridge + Lambda g6 capacity guarantees के लिए

📚 Reading​

  • Designing Data-Intensive Applications (re-read, Chapter 7: Transactions)
  • NVIDIA Blackwell architecture whitepapers
  • The Staff Engineer's Path by Tanya Reilly

🎤 Speaking​

📫 Open To​

  • Senior/Staff/Principal roles: DevOps, Platform, MLOps, Cloud Architecture
  • Consulting: GPU inference platforms, EKS hardening, cloud cost optimization
  • Conversations: GPU scheduling, FinOps, Kubernetes security, observability

Last updated: October 2026. यह page monthly change होता है।