What actually happens when you build with language models.
Long-form notes from evenings spent running models on a single GPU, breaking retrieval pipelines and learning to distrust my own benchmarks. No hype, no newsletters, no trackers.
Evals: how not to fool yourself
Your new prompt scored 4% better. Is that real? A practical guide to golden sets, LLM-as-judge biases, variance, and the habit of looking at the actual outputs.
Retrieval-augmented generation without the hype
Most "the model hallucinated" bugs in RAG are retrieval bugs. Chunking, hybrid search, reranking — and measuring each step separately.
Running a 14B model on a 12 GB GPU
The back-of-the-envelope memory math for weights and KV cache, what 4-bit quantization really costs you, and where the VRAM actually goes.
Tokenizers are weirder than you think
Why models can't count letters, why leading spaces matter, and why the same sentence costs three times more in some languages.