Notes from the den.
The same posts as a plain list, with filters and paging. The default view is the shelf. Want to write one? Get in touch.
Making Research Papers Sound Less AI-Written
No single word gives the writing away. Readers notice the pattern, and the fix is stronger scholarship rather than cosmetic humanization.
Md Sultanul Islam Ovi
Great on Training Data, Useless Everywhere Else
Overfitting, distribution shift, and weak evaluation, untangled. Three failures that look the same and need different fixes.
Jamal Hossain
Making a Model Smaller: Quantization and Pruning for Beginners
Why a model that fits in memory and a model that runs well are two different problems, and how to close the gap between them.
S.M Seefat Alam
Running Applications as Windows Services: Architecture, Implementations, and Best Practices
What the SCM expects from your process, why anything with a window hangs, and four ways to register a project.
Miraj Uddin Chowdhury
Data to Power: AI, the Invisible Compute Economy, and the Rise of Silicon and Energy Systems
A token is a unit of work, not a unit of text. Follow one down and it ends at a semiconductor fab, then a power grid.
I Built a RAG System That Skips the LLM Whenever It Can. Here's Why That Was the Right Call
Most catalog questions are lookups, not reasoning tasks. A deterministic path answers them in under 10ms, and generation becomes the last resort.
Ridam Roy
Unsloth: How a Rewritten Backward Pass Made Fine-Tuning LLMs a Free-Tier-GPU Problem
Hand-derived gradients and Triton kernels, and why they buy more than a bigger GPU budget does.
What Is CUDA, and How to Run LLMs Locally: A Beginner's Guide
What CUDA actually does, how to verify yours works, and three ways to load a model onto your own GPU.
Playwright vs. Cloudflare Turnstile: What Bot Detection Actually Looks For
Four layers, from the TLS handshake to cursor trajectories, each weak alone and hard to satisfy at once.
Working With Low-Resource and Multilingual Text
Tokenizers that fragment meaning, benchmarks that measure the wrong thing, and the dialect choices nobody writes down.
Muhsina Tarannum Munfa
Imposter Syndrome in a Room Full of Smart People
Competence doesn't cancel out doubt. It just gets better at hiding in more impressive rooms.
Building Your First Dataset, and the Responsibility That Comes With It
What you collected and what you called noise decides most of what a model can ever get right.
9 RAG Architectures You Should Actually Know in 2026
Naive RAG, reranking, hybrid search, GraphRAG, agentic and multimodal retrieval: what each one fixes, what it costs, and how to pick.
Writing Research Code That Isn't a Single 800-Line Notebook
A five-stage pipeline is a directed acyclic graph, not a linear script. What belongs in a script, and what belongs in a notebook.
Ann Naser Nabil
Parsing PDFs Is Broken: How Docling and PixelRAG Are Fixing Unstructured Data
When your RAG system hallucinates numbers, the problem is usually the ingestion pipeline, not the model.
My Model Won't Learn: A Debugging Checklist
2.30 at step 0, 2.30 at step 500. The order to check things in, starting with the test that splits the problem in half.
Md. Tanvir Mahamud Himel
Config-Driven Experiments, or How I Stopped Editing Code to Change a Learning Rate
Hyperparameters hardcoded across four scripts, and a paper that reported two different F1 values for the same model.
The Writing Mistakes in Almost Every First Paper
Mechanics, structure, framing, evidence, and compliance: the five places a manuscript loses its reviewer.
Experiment Tracking Before You Have 200 Runs You Can't Tell Apart
Four hundred runs, three tools, and no way to answer which one produced the numbers in the paper.