Blog

• vLLM vs SGLang on a Single A10G: What 16 Models Actually Showed — How I stopped trusting tokens-per-second charts and started measuring the things that page you at 3 AM. 16 models, 5 scenarios, both engines, one 24 GB A10G, and 533 result files later: there is no winner, the two biggest gaps trace back to a single execution flag, and the results worth acting on are the p99 blowup and the goodput inversion that no throughput chart shows.
• Mamba vs Transformers: What I Learned Chasing a Context Window — Six years ago a chatbot I was building fell apart on long conversations, and I could not explain why. This is the answer I finally chased down: the quadratic cost of attention, what selective state space models do about it, a GPT-2 profiling run that hits the wall on a free Colab T4, and why hybrid Mamba-Transformer models are the architecture that actually won.
• Grow a New Tree with Git Worktrees — A calm way to ship a hotfix without wrecking your feature work. You open a second folder from the same project, tied to a different branch. Your feature stays exactly where you left it. The production fix lives in its own clean space. Fix it, push it, delete the folder, and step right back into your feature like nothing happened.
• TrustMed AI: Building a RAG-Powered Healthcare Chatbot on AWS — A deep-dive into building an AI-powered chatbot for Type II Diabetes & Cardiovascular Disease using Retrieval-Augmented Generation (RAG) with Amazon Bedrock, OpenSearch, and LangChain.
• Implementation of a Serverless Data Lake for Scalable Real Estate Analytics — A deep dive into re-architecting a real estate platform using a Serverless AWS Data Lake approach, achieving sub-3-second query latency while handling 10M+ logs without provisioning EC2 instances.
• Cold Starts in AWS Lambda and How to Mitigate Them — Cold starts were the hidden tax on my serverless bet. Here’s the short, practical playbook for warm starts, lower latency, and smaller bills.
• From Spikes to Steady Throughput Using SQS and Lambda — Architecture workflow of the ETL pipeline