Occasional notes on AI engineering, on-device ML and the modern web.
Why an unquantized 12B won't run on a gaming GPU, how QLoRA and 4-bit quantization make it fit, and how a fine-tuned draft model makes it 1.7x faster, fast enough for a live voice call.
How md2docx went from a VPS container to serverless on AWS for ~$0 — without changing a line of app code, plus the three sharp edges nobody warns you about.
An assistant that auto-replies to email can do real damage. The boring safety layer is what makes it shippable.
Word, Excel and PowerPoint are ZIP archives of XML. Here's how to change the text and keep every font, table and chart intact.
Fast local classification you can trust, plus an LLM that justifies each call with quotes — without letting it touch the verdict.
Cloud models are convenient until the data is confidential. Why I keep shipping apps where the model runs entirely on your machine.
Two people, one slot, the same millisecond. Application-level checks lose this race — Postgres can win it.
Most RAG demos fall over in production. Here is the small set of things that actually made mine reliable — and measurable.