Mahamudul Hasan — Writing

Mahamudul Hasan — Writing https://mahamudul-hasan-cse.github.io/blog/ Writing on offline AI systems, RAG pipelines, local LLMs, computer vision, and the gap between a working demo and a working product. en-us Fri, 20 Mar 2026 00:00:00 +0600 mahamudulh417@gmail.com (Md. Mahamudul Hasan) mahamudulh417@gmail.com (Md. Mahamudul Hasan) My Model Hit 85% Accuracy. Here's What That Number Is Hiding. https://mahamudul-hasan-cse.github.io/blog/ml-mental-health-model.html https://mahamudul-hasan-cse.github.io/blog/ml-mental-health-model.html Fri, 20 Mar 2026 00:00:00 +0600 Class imbalance, the asymmetry of false negatives in mental health prediction, and why the metric you optimize determines who gets missed. What I should have written in the README. Machine Learning The Web Is Full of Garbage (For LLMs): Why I Built My Own Crawler https://mahamudul-hasan-cse.github.io/blog/crawl4ai-llm-web-scraping.html https://mahamudul-hasan-cse.github.io/blog/crawl4ai-llm-web-scraping.html Sun, 02 Mar 2026 00:00:00 +0600 Raw web scraping gives you text. LLMs need structure, metadata, and clean extraction ratios. Why standard scrapers fail for AI training data, and what Crawl4AI does differently. Developer Tools Building an AI for Farmers Who Speak Bengali, Farm Without Internet, and Don't Have Time for Demos https://mahamudul-hasan-cse.github.io/blog/building-krishibot-bilingual-ai.html https://mahamudul-hasan-cse.github.io/blog/building-krishibot-bilingual-ai.html Sat, 14 Feb 2026 00:00:00 +0600 What KrishiBot actually taught me about bilingual NLP, the Banglish input problem, offline crop disease detection, and the response length insight that mattered more than any model improvement. LLMs Running RAG Completely Offline: What Nobody Tells You Until It Fails https://mahamudul-hasan-cse.github.io/blog/offline-rag-no-api.html https://mahamudul-hasan-cse.github.io/blog/offline-rag-no-api.html Sat, 10 Jan 2026 00:00:00 +0600 I built two RAG systems that run entirely on CPU with no internet. Here's everything that broke — chunking strategies for equations, FAISS persistence quirks, LLM hallucination on domain-specific content, and the one config change that cut latency in half. RAG Hands-Free Mouse Control With a $0 Hardware Budget https://mahamudul-hasan-cse.github.io/blog/hands-free-mouse-webcam.html https://mahamudul-hasan-cse.github.io/blog/hands-free-mouse-webcam.html Sun, 28 Dec 2025 00:00:00 +0600 The EAR algorithm, exponential smoothing, and why distinguishing an intentional blink from a reflex is harder than it looks. How I shipped a hands-free mouse controller that runs on any laptop with a webcam. Computer Vision Why I Stopped Using Cloud LLMs and Built My Own https://mahamudul-hasan-cse.github.io/blog/self-hosted-llm-localmind.html https://mahamudul-hasan-cse.github.io/blog/self-hosted-llm-localmind.html Mon, 15 Dec 2025 00:00:00 +0600 The real cost of cloud AI isn't the money — it's your data leaving your machine. What I built with LocalMind AI, what surprised me most, and the architecture decisions that made streaming feel as fast as GPT-4. LLMs