Part 10: Inside NVIDIA Newton: Open-Source Differentiable Physics for Generalist Robotics
Explore NVIDIA Newton, the open-source differentiable physics engine co-developed with Google DeepMind and Disney Research for robot learning.
My Blog
Reflections on software engineering, pipeline automation, graphics, and system reliability.
Filter by Date:
Part 9: Demystifying Autonomous Vehicles: The 3-Computer Architecture, SAE Autonomy Levels, and the Sensor Fusion Triad
Explore the 3-computer AV architecture, SAE Levels 0 to 5, the Perception-Planning-Control triad, and multimodal Camera-Radar-LiDAR sensor fusion.
Part 8: Silicon at the Edge: NVIDIA Jetson Thor Architecture & Isaac ROS Acceleration
Explore NVIDIA Jetson Thor edge computing, Blackwell 800 TFLOPS architecture, NITROS zero-copy IPC, and Isaac ROS sub-50ms humanoid reflexes.
Part 7: From Simulation to Streets: NVIDIA DRIVE & Alpamayo Autonomous Vehicle Architecture
Explore NVIDIA DRIVE Thor, Alpamayo AV foundation models, surround-view Bird's-Eye-View (BEV) transformer fusion, and ASIL-D safety-critical redundancy.
Part 6: Inside Project GR00T: Vision-Language-Action (VLA) Tokenization & Diffusion Action Heads
A deep architectural dissection of NVIDIA Project GR00T: multimodal VLA tokenization, transformer cross-attention backbones, and diffusion policy action heads for humanoid robotics.
Part 5: Scaling Physics with Isaac Sim & Omniverse Replicator: GPU Dynamics, Synthetic Sensors, and Domain Randomization
An in-depth engineering guide to NVIDIA Isaac Sim and Omniverse Replicator: GPU-accelerated PhysX 5 dynamics, synthetic sensor pipelines, and automated domain randomization.
Part 4: Demystifying OpenUSD: Architecture, Composition Arcs, usdview, and Simulation Assets
A comprehensive guide to OpenUSD (Universal Scene Description): hierarchical scene graphs, LIVRPS composition arcs, step-by-step usdview visualization, and SimReady assets.
Part 3: Unlocking NVIDIA Omniverse: Architecture, OpenUSD, RTX Rendering, and the Industrial Metaverse Ecosystem
A comprehensive architectural deep-dive into NVIDIA Omniverse: OpenUSD scene graphs, Nucleus live-sync collaboration, RTX real-time ray tracing, and industrial digital twin ecosystems.
Part 2: Inside NVIDIA Cosmos: World Foundation Models for Physical Commonsense & Video Trajectories
A deep architectural dive into NVIDIA Cosmos world foundation models: Mixture-of-Transformers (MoT), continuous spatiotemporal latent tokenizers, and physics-aware trajectory generation.
Part 1: Unpacking the NVIDIA Physical AI Data Factory (PAIDF) Stack
An architectural deep-dive into NVIDIA's Physical AI Data Factory (PAIDF) stack: synthetic data generation with Cosmos & Isaac Sim, VLA foundation models, and edge runtime orchestration.
NVIDIA Nemotron 3.5 Lightning Deep-Dive: 30B MoE Architecture, 3B Active Params, and Local Ollama Execution
How NVIDIA's Nemotron 3.5 Lightning achieves 3B active parameter speed with 30B MoE capacity, 1M context window, and native Ollama local agent execution.
Frontier MoE Deep-Dive: Analyzing Alibaba's Qwen 3.8 Flagship Architecture, Performance, and Token Pricing
How does Alibaba's Qwen 3.8 flagship model achieve GPT-5 class reasoning at $0.30 per 1M tokens? Explore 512-expert sparse MoE routing, Multi-Head Latent Attention (MLA), performance benchmarks, and token economics.
Part 9: The Evolutionary Arc of Computer Vision: From LeNet-5 and ResNet to ConvNeXt and 3D Video Models
How did computer vision evolve from reading bank check zip codes in 1998 to 3D video spatial physics in physical AI and robotics?
Part 8: Generative Adversarial Networks (GANs): The Counterfeiter vs. Detective Minimax Game
Step into the zero-sum game of generative AI—how a counterfeiter Generator and detective Discriminator compete to reach Nash Equilibrium and synthesize photorealistic images.
Part 7: The Attention Memory Bottleneck: From Self-Attention Basics to MHA, GQA, and DeepSeek's MLA
How does DeepSeek-V3 run 128K context windows with 96.5% less VRAM? Demystify Key-Value Cache growth, MHA, GQA, and low-rank Multi-Head Latent Attention (MLA).
Part 6: Why Deep Networks Die: Weight Initialization (He/Xavier), LayerNorm, and Residual Connections
Why can't you train a 100-layer neural network without it blowing up or learning nothing, and how do ResNet shortcuts and LayerNorm keep it alive?
Part 5: Inside the Learning Engine: Forward Pass, Backpropagation, and Dynamic Autograd
How do neural networks actually learn—how do predictions flow forward, loss errors walk backward via the chain rule, and autograd record every operation?
Part 4: Demystifying Activation Functions: Why Neural Networks Need Non-Linearity, Types, and Real-World Use Cases
Why are linear models completely useless for complex real-world data, and how do activation functions like ReLU, GELU, and SwiGLU warp space?
Part 3: The Transformer Revolution: How Self-Attention and Q K^T V Solved the GPU Parallelization Bottleneck
How did Query-Key-Value self-attention and matrix parallelization eliminate sequential GPU bottlenecks to power ChatGPT?
Part 2: Why LSTMs Were Needed: Conquering RNN Amnesia, Memory Conveyor Belts, and Gated Doors
Why do standard RNNs suffer from total amnesia on long sequences, and how did LSTMs solve it with memory conveyor belts and gated doors?
Part 1: Demystifying Neural Networks: From Simple Perceptrons to Deep Neural Networks (DNNs), CNNs, and RNNs
How did AI evolve from single-layer perceptrons to deep feedforward networks, 2D spatial CNN filters, and sequential RNN loops?
Google DeepMind's Gemini Robotics ER 2: The High-Level Brain for Physical AI and Multi-Robot Collaboration
An architectural deep-dive into Google DeepMind's Gemini Robotics ER 2 announcement—exploring real-time video streaming, high-level reasoning vs low-level VLA execution, multi-robot team orchestration, and temporal moment-finding.
Demystifying LoRA (Low-Rank Adaptation): From Training Efficiency to Multi-Adapter Inference
A comprehensive developer guide to Low-Rank Adaptation (LoRA). Exploring why LoRA is used, matrix decomposition math, VRAM reductions during training, weight merging vs multi-adapter serving during inference, QLoRA, DoRA, and runnable Python simulations.
Deep-Dive: SGLang v0.5.16 Architecture and High-Throughput Inference Comparison
An architectural deep-dive into SGLang v0.5.16. Analyzing RadixAttention KV cache reuse, compressed FSM constrained decoding, Torch Compile CUDA graph optimizations, and multi-engine benchmarks.
Understanding Mixture-of-Experts (MoE): From Specialist Clinics to Kimi K3's 896-Expert Router
A developer-friendly guide to Mixture-of-Experts (MoE) architectures. Exploring the specialist clinic analogy, total vs active parameters, Top-K gating networks, expert collapse traps, and how Kimi K3 scales to 896 micro-experts.
Hosting Moonshot AI's Kimi K3 Open Weights with vLLM: High-Throughput Serving at Scale
A comprehensive developer guide to hosting Moonshot AI's Kimi K3 open weights on vLLM. Exploring MXFP4 MoE serving, KDA hybrid prefix caching, DSpark speculative decoding (370 tok/s), and NVIDIA/AMD multi-GPU cluster recipes.
Physical AI Models: Grounding Intelligence in Space, Physics, and Robotics
A technical deep-dive into Physical AI models. Exploring spatial understanding, 3D geometry priors, physics simulation engines, embodied VLA control loops, and robotics integration.
The DeepSeek Architectural Inflection Point: From MLA to Emergence of Open-Weight Reasoning
An architectural deep-dive into DeepSeek's evolution from DeepSeek-LLM to R1. Exploring Multi-Head Latent Attention (MLA), DeepSeekMoE, MTP, pure RL emergence, and comparisons with contemporary frontier models.
Scale and Performance: Serving LLMs with vLLM and llm-d
A comprehensive developer guide to scaling LLM serving using vLLM and llm-d. Explore PagedAttention, continuous batching, disaggregated prefill/decode, and Kubernetes deployment scripts.
Under the Hood of Moonshot AI's Kimi K3: The Architecture of 3-Trillion Parameter Thinking Models
A comprehensive developer guide to Moonshot AI's Kimi K3. Comparing K3, K2.7 Code, K2.6, and K2.5 across Preserved Thinking, reasoning effort, 1M context, API quickstart, organizational best practices, and prompt engineering.
Anthropic's Claude Opus 5: Frontier Reasoning, Benchmarks, and Prompt Engineering
A deep-dive into Anthropic's flagship Claude Opus 5 release. Comparing Opus 3 vs. Opus 3.5 vs. Opus 5, analyzing benchmarks, and mastering official prompt engineering techniques.
The Architectural Spectrum of World Foundation Models: Renderers, State Simulators, and Action Planners
An engineering breakdown of World Foundation Models—moving beyond text tokens into spatial intelligence, persistent 3D state representations, and perception-action control loops.
Google's Gemini 3 Family: The Comprehensive Developer Guide and Model Comparison
An in-depth guide to Google's newly expanded Gemini 3 model family—including Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber, and 3.1 Pro. Learn how they compare on latency, cost, and capabilities.
Connecting Telegram to OpenClaw: A Complete Step-by-Step Guide
Learn how to connect OpenClaw to Telegram from scratch. A beginner-friendly step-by-step tutorial using BotFather to run your AI butler locally.
LangChain vs. LangGraph: Moving from Chains to Cyclic State Graphs
Why linear chains fail at complex agentic workflows. Learn how LangGraph models agents as state machines using persistent schemas, nodes, and cyclic edges.
vLLM vs. llama.cpp: Which is the Real Production King?
A deep-dive architectural comparison between vLLM and llama.cpp. Compare throughput, memory footprints, and choose the right engine for your workload.
Thinking Machines' Inkling: Under the Hood of the 975B Parameter Open Multimodal MoE
A deep dive into Mira Murati's new open-weights multimodal model, Inkling. Analyze the 975B MoE architecture, active parameter routing, and local serving.
Token Economics, LLM Gateways, and Router9
Understand the financial side of LLM serving. Calculate cost-per-token, setup open-source LLM gateways, and implement intelligent routers like Router9.
Inference Optimizations: Speeding up Prefill and Decode
Explore advanced LLM serving optimizations. Learn how FlashAttention, chunked prefill, speculative decoding, and KV cache eviction speed up TTFT and inter-token latency.
OpenClaw in Action: Connecting WhatsApp to Automated Workflows
A step-by-step developer guide to configuring OpenClaw's WhatsApp gateway, setting up secure session authorization, and building automated task sync skills.
The Self-Hosted AI Butler: Modular Assistance with OpenClaw
Learn how to build and host your own AI assistant with OpenClaw. Configure the ClawHub modular skill store, write SKILL.md files, and connect to chat platforms.
Nous Research's Hermes Agent: Under the Hood
Deep-dive into the configuration, CLI commands, and skill architecture of Nous Research's Hermes Agent. Learn to manage tool gateways and design custom skills.
The Landscape of Agentic AI: From Single-Agent Scripts to Multi-Agent Networks
Start the Autonomous AI Agents Series. Learn about the ReAct pattern, single-agent context decay, and multi-agent coordination graph topologies.
The Landscape of LLM Inference Engines: Open Source vs. Enterprise
A comprehensive developer guide comparing vLLM, TensorRT-LLM, TGI, SGLang, and llama.cpp. Learn about PagedAttention, RadixAttention, and GGUF.
Understanding the KV Cache: The VRAM Bottleneck of LLM Serving
Why do LLM serving nodes run out of VRAM? Explore the mechanics, calculations, and capacity bottlenecks of the Key-Value (KV) Cache.
The Two Pillars of LLM Inference: Prefill vs. Decode
Explore the two execution phases of LLM inference: compute-bound prefill vs. memory-bandwidth-bound decode. Understand arithmetic intensity and the role of the KV cache.
Basics of AI Inference: Demystifying Latency, Throughput, and Serving
Start the AI Inference Deep-Dive Series. Learn the fundamentals of inference vs training, key performance metrics, and the systems engineering behind serving.
The Model Taxonomy: LLMs, Vision Models, VLAs, and Diffusion
A developer's guide to the modern AI model taxonomy. Understand the architecture, modalities, and performance profiles of LLMs, ViTs, VLAs, and Diffusion.
Training vs. Inference Lifecycle: A Developer's Guide to Weights, Backpropagation, and Serving
A step-by-step developer's view of the AI model lifecycle—from raw weights and backpropagation during training to freezing weights and serving inference.
What is a Model Weight? Demystifying Tensors, Matrices, and File Formats
What actually happens when you download a model from Hugging Face? A developer's guide to tensors, weights, biases, and serialization formats like Safetensors and GGUF.
Claude Code Custom Plugins: Build and Host Your Own Marketplace
Learn how to build, structure, and host a custom, decentralized Claude Code plugin marketplace using a simple GitHub repository.
Running Local LLMs: Ollama vs. vLLM
A comprehensive guide to running Large Language Models locally on Windows, macOS, and Linux using Ollama and vLLM, including pros, cons, and performance tuning.
Anthropic's Claude Model Family: Specs, Pros, Cons, and Use Cases
Compare Anthropic's Claude models—Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5, and Claude Science. Learn their pros, cons, and the best use cases.
Anthropic's Mid-2026 Wave: Claude Sonnet 5, Claude Science, and Fable 5 Redeployment
A summary of Anthropic's major announcements: the launch of Claude Sonnet 5, the new Claude Science AI workbench, and the global redeployment of Fable 5.
Claude Code Custom Skills: Design Methodology and Workspace Personas
Learn how to design, structure, and implement persistent, context-aware custom agent skills in Claude Code to automate code reviews and git workflows.
Claude Code Basics: Commands, Subagents, and Memory Layers
Learn the difference between slash commands and file context mentions, configure global vs local memory, and understand how custom skills and subagents operate in Claude Code.
Claude Code Customization: CLAUDE.md, AGENTS.md, and SKILLS.md
Master project-level customization for Anthropic's Claude Code by writing rules and guidelines in CLAUDE.md, AGENTS.md, and custom SKILLS.md.
Setting up Claude Code: The Ultimate Terminal AI Pair Programmer
A step-by-step setup guide for installing, configuring, and authenticating Anthropic's Claude Code CLI across macOS, Windows, and Linux.
LLMs, Agents, and Harnesses: Demystifying Claude Code
Confused about the difference between Claude Code and Claude Opus? A deep dive into the architecture of modern agentic coding tools—from Large Language Models (LLMs) to Autonomous Agents and Execution Harnesses.
Custom Model Endpoints: Hooking up Local & Enterprise Models with Antigravity CLI
Learn how to configure the Google Antigravity CLI (agy) to run with local LLMs (Ollama, vLLM) or private enterprise endpoints for maximum privacy and cost efficiency.
Going Async: Background Tasks, Timers, and Scheduling in Antigravity CLI
A deep dive into managing long-running tasks, using asynchronous tools, setting timers, and configuring recurring jobs with the Google Antigravity CLI.
Extending Antigravity CLI: Building Custom Skills and Project Rules
Learn how to supercharge the Google Antigravity CLI (agy) by building custom agent skills, project-scoped rules, and custom scripts to automate repository-specific workflows.
Setting up Google Antigravity CLI: The AI-First Terminal Assistant
A comprehensive guide on installing and configuring Google Antigravity CLI (agy) on Mac, Windows, and Linux, complete with TUI commands, settings.json customization, and real-world usage examples.
Python 3.15: A First Look at the Coolest New Features
A deep dive into the most exciting developer features in the Python 3.15 pre-release, including explicit lazy imports, frozendict, built-in sentinels, unpacking in comprehensions, and the Tachyon profiler.
SLURM Demo on AWS Ubuntu EC2 instance
Demo of slurm usage on a single instance of Ubuntu 24.04 EC2 instances on AWS
SLURM on WSL
Setting up SLURM on WSL
Introduction to SLURM
Simple Linux Utility for Resource Management (SLURM)
How to find a linux machine is a VM (Virtual Machine) or a Bare Metal
If you can SSH into a linux machine and want to find out if its baremetal or Virtual Machine
Storing Github access token in git credential store
Using git credentials store the github access token to avoid the re-prompting of username and pwd
Token generation for Registering Self Hosted Github Runner via REST API
Explains how to generate a token using github API to be used in turn with Github self hosted runner registration
Setting up a Self Hosted Github Runner
Explains how to setup a Github self hosted runner and register
Managing the NodeJS versions on Windows
Node Version Manager (nvm) helps in managing multiple NodeJS versions
Customizing the Powershell terminal with oh-my-posh
Instructions on customizing the terminal in powershell with oh-my-posh and winget
File permissions on Windows - chmod 400 in Powershell
Powershell equivalent of chmod 400 for Windows files
Github login using access token via command line
Logging in using github access token (no more passwords)
Adding Google Analytics to NuxtJS app
Adding Google Analytics GA4 property to NuxtJS App
Productive Taskbar Settings missing in Windows 11
Very useful Taskbar Settings goes missing in Windows 11.
Secureboot + Ubuntu + VirtualBox Signing kernel modules
Set of steps required for dealing with secureboot on Ubuntu where VirutalBox service has issues
Fixing the postfix error dpkg
Steps to fix the postfix error happening during apt upgrade ubuntu.
Running a react app on Local Kubernetes cluster on Windows 10
Process and steps for running react app on local k8s cluster using minikube on windows 10
Gatsby site hosted on AWS Amplify redirecting to homepage always
Using the rewrites and redirects on AWS Amplify for the depolyed personal website
Connecting AWS Amplify for deployment of website
Explains how to connect the gatsby website hosted on github to AWS Amplify for deployment
Evolution of this website
will be capturing the evolution of my website chronologically (latest first)
Making of "Tree Story" Animated Short
Behind the scenes of the Animated Short "Tree Story"
Profiling & Visualization Tools in Python - Part 1
Profiling & Visualization Tools in Python using cProfile
Missing permissions issue linking google analytics and adsense
Problem of linking google analytics and adsense
Adding Google Analytics to personal website
How I added google analytics to my personal website written in Gatsby
Chrome Remote Desktop on Fedora
How to install Chrome Remote Desktop on Fedora
Building grpc whl with Maya 2019 on Windows
My stuggle with building a grpc python library with Maya 2019 on Windows
Tech Stack of the website
Brief about the tech stack used for this website
Deleting linux from dual boot
Steps for safely removing linux from dual boot.