https://kolm.ai/articles Engineering essays from Kolmogorov: AI compilers, verified inference, on-device HIPAA, the .kolm artifact format, and recipe-driven speculative decoding. en-us Wed, 07 May 2026 12:00:00 GMT https://kolm.ai/articles/ai-compiler https://kolm.ai/articles/ai-compiler Thu, 07 May 2026 12:00:00 GMT An AI compiler is the missing build step between a frontier model API and a local artifact. This essay explains what it does, why it changes the unit economics of AI, and how to compile your first .kolm in five minutes. AI compiler distillation on-device https://kolm.ai/articles/k-sample-verified-inference https://kolm.ai/articles/k-sample-verified-inference Thu, 07 May 2026 12:00:00 GMT K-sample verified inference produces cryptographically auditable model outputs without the cost of zero-knowledge proofs. The same mechanism powers every label inside a .kolm artifact. verified inference audit receipts https://kolm.ai/articles/hipaa-on-device https://kolm.ai/articles/hipaa-on-device Thu, 07 May 2026 12:00:00 GMT A practical architecture playbook for keeping protected health information local when evaluating AI assistants for healthcare workflows. HIPAA clinical on-device https://kolm.ai/articles/kolm-file-format https://kolm.ai/articles/kolm-file-format Thu, 07 May 2026 12:00:00 GMT A field-by-field walkthrough of the .kolm file format: base model, LoRA adapter, recipe pack, recall index, verifier, held-out tests, manifest, signature. The container behind portable AI artifacts. .kolm spec format https://kolm.ai/articles/speculative-decoding-recipes https://kolm.ai/articles/speculative-decoding-recipes Thu, 07 May 2026 12:00:00 GMT A draft model is the standard way to do speculative decoding. A deterministic recipe pack is faster, smaller, free at runtime, and verifiably correct. Here is how to use one to cut your local LLM bill to zero. speculative decoding drafts runtime