Thoughts, insights, and experiences from my journey in technology and research.
In a pull-based, available-under-partition system the same task deliberately runs more than once, so how does a result land exactly once? A two-phase staging commit copies a staged object into an identity-derived final key so duplicates collapse to one, a Raft-agreed decision fences split-brain, and the one irreducible side-effect window is made safe with an idempotency key rather than wished away.
Two scheduling problems that share a backbone. Bounded-grant leases let a second worker help finish a hot shard while proving nothing is ever processed twice, and a load-shedding admission controller turns each tenant into an Erlang-B loss system whose overload behavior you can predict on paper instead of queueing work until latency collapses.
A Deep Dive into How C++ Evolved to Make Memory Safe, Efficient, and Expressive
C++ is a powerful, high-performance language that fully supports Object-Oriented Programming (OOP). Unlike procedural programming, which focuses on functions and logic, OOP focuses on objects that combine data and behavior.
This project implements a production-grade distributed inference system in C++ that uses consistent hashing to route requests to worker nodes for optimal cache utilization, circuit breakers to handle failures gracefully, LRU caching to avoid redundant computation, and dynamic batching to maximize hardware efficiency during model inference. The architecture combines a central gateway that manages request routing and fault tolerance with multiple worker nodes that each run ONNX Runtime engines wrapped in caching and batching layers.
This project implements a production-grade distributed inference system in C++ that uses consistent hashing to route requests to worker nodes for optimal cache utilization, circuit breakers to handle failures gracefully, LRU caching to avoid redundant computation, and dynamic batching to maximize hardware efficiency during model inference. The architecture combines a central gateway that manages request routing and fault tolerance with multiple worker nodes that each run ONNX Runtime engines wrapped in caching and batching layers.
Scaling machine learning models for production requires more than just a powerful GPU. It requires a distributed architecture capable of handling high throughput while maintaining low latency. This post explores the implementation of a distributed inference system utilizing consistent hashing for load balancing and dynamic request batching for efficiency.
Matrix algebra forms the bedrock of many modern applications in computer science, data science, physics, and engineering. From transforming images to training machine learning models, matrices are everywhere—quietly powering the math behind the scenes. This blog is an extension to the previous blog on Matrix Algebra covering more advanced concepts.
Matrix algebra forms the bedrock of many modern applications in computer science, data science, physics, and engineering. From transforming images to training machine learning models, matrices are everywhere—quietly powering the math behind the scenes. In this blog, we’ll break down the basics of matrix algebra. Whether you're just getting started or brushing up, this guide will help you build a strong foundation in understanding how matrices work and why they're so essential.
Neural networks have revolutionized how machines learn from data—powering everything from voice assistants to medical diagnosis tools. But beneath the buzzwords and real-world applications lies a fascinating world of math that makes it all work. In this blog, we’ll dive into the mathematical foundations of neural networks—covering concepts like weighted sums, activation functions, backpropagation, and gradient descent—while also exploring how these powerful models are used across domains like computer vision, natural language processing, healthcare, and beyond.
This project tackles plant disease detection using a multi-class classification approach with CNNs, leveraging EfficientNet for its high accuracy and efficiency on limited hardware. To further optimize performance, a two-stage pipeline was introduced: first identifying the plant species, then classifying the disease using species-specific models. This hierarchical method not only reduced computational load but also improved accuracy by narrowing the classification scope in the second stage.
Growth often feels invisible—until it isn’t. Inspired by the quiet power of bamboo, this blog explores how foundational work in learning and problem-solving, especially in machine learning and AI, mirrors nature’s most patient plant. This post dives into how embracing the “root-growing” phase can transform how we approach challenges, projects, and growth itself.