Writing

Field notes from production.

Everything here was learned the expensive way — on live traffic, real GPUs, and customers who don't forgive a hallucinated return policy.

Jun 14, 2026

RAG that refuses to guess

An 8B model asked about a refund policy it cannot find will invent one, and sound certain. Here is the graph that takes that decision away from it.

RAGLangGraphProduction AI
11 min
May 20, 2026

Arabic search is three search problems in a trench coat

MSA, Iraqi dialect and Franco-Arabic do not tokenize alike, spell alike, or fail alike. What it takes to build retrieval that survives all three.

SearchArabic NLPQdrant
9 min
Apr 18, 2026

The flywheel: turning live traffic into better models

The first LoRA adapter is easy. The second one is the real question: where does training data come from when nobody has time to annotate?

Fine-tuningLoRAMLOps
12 min
Mar 10, 2026

Two agents, one T4: self-hosting on a budget

A vision-language model, two LoRA adapters and an embedding model on 16 GB of 2018 silicon. The flags that make it fit, and the two that quietly destroy it.

vLLMInferenceQuantization
10 min
Feb 8, 2026

Shipping AI agents into Instagram DMs

One private reply per comment. Webhooks that retry if you are slow. A self-loop that rate-limits you into a 500. The platform is the spec.

Meta APIAgentsWebhooks
9 min