Writing

Field notes on AI infrastructure, LLM inference on real hardware, systems programming, and the schemas nobody warns you about.

j/k to navigate · ↩ to open

Aug 25, 2026Constrained Generation: Making Hallucinated SQL Structurally ImpossibleThe most reliable fix for Text-to-SQL hallucination isn't a better prompt, it's making wrong answers unrepresentable.

Jul 15, 2026Seccomp, Namespaces, and the Case for Kernel-Level Agent SandboxesPrompt-level guardrails are suggestions. Linux namespaces and seccomp BPF are physics, here's the mechanism behind kernel-level agent sandboxing.

May 20, 2026The 8GB VRAM Budgeting Playbook: Two Models, One GPUQuantization tables, load order, KV-cache budgeting, and the runtime choices that decide whether your inference box survives the week.

Apr 2, 2026RAG That Survives Production: Grounding Patterns Beyond the DemoChunk size is the least interesting decision. Corpus quality, citation discipline, and failure semantics decide whether RAG holds up.