Blog

A technical introduction to how LLMs work, how they are trained, and what fine-tuning really does. Built from 18 papers read together, organised around one idea: post-training surfaces knowledge, it does not add it.

A working reference for retrieval-augmented generation: how to find the stage that is failing, what to measure, and when RAG is the wrong tool. Retrieval, chunking, metadata, hybrid search, reranking, evaluation, observability, permissions, freshness, cost, and the alternatives.

SemRec: a 940MB fine-tuned Qwen2.5-Coder model that recovers clean semantics from obfuscated Python. 93.36% overall, 11/11 on competition-grade obfuscation.

A 25M parameter model, a strict little DSL, and a browser toy that runs locally with WebGPU or WASM.

How I built a self-hosted SMS gateway using a spare Android phone, a custom FastAPI relay, and Tailscale for secure, zero-cost text messaging from my AI agent.

AI-powered K-12 educational platform. Generates curriculum-aligned worksheets, quizzes, and info sheets. Live at https://educated.space

How I used WhisperX forced alignment to fix inconsistent word timings in a Quran reading app.

Building a full-stack Quran exploration app with React 19, FastAPI, and a hybrid database design.

A step-by-step guide to setting up your Steam Deck to run local Large Language Models (LLMs) using llama.cpp and a Distrobox container.

How I fine-tuned Llama 3.2 3B to process transcripts locally, achieving better results than models 5-8x larger through specialized training.

Designing and building a distraction-free Android app in Kotlin/Compose with low-latency audio and a gentle 'Find it' game.

Turning a 1B LLM into a deranged cosmic entity running on CPU only.