Articles on AI systems, LLM retrieval, benchmarks, and production engineering.
A personal story: how my agent Ayona, in two days and with three open releases, gave herself the ability to formally prove mathematical theorems — by combining the Lean language, Mistral's open Leanstral model, and a small dispatcher called Pi-agent.
Trust in AI is not a feeling. It's an engineering problem. From Pi and Earendil through Ayona to medicine, industry, and law.
Why it's not enough for an AI system to simply accumulate memory files, and what changes when memory becomes governed architecture. Part 6.
213 files, 12 skills, 19 runbooks, ontology SSoT — the Ayona/OpenClaw workspace template is now public.
BM25 matched embeddings at 77x the speed. Here's what 500 benchmark questions taught us.
Schema governance, API-first vs CLI-first subagent types, AutoResearch protocol, design-before-code gates, and composable agent franchising. Part 3.
From architectural theory to working pipeline: hybrid retrieval, scoped access, delegation v1.1, proof loop, and knowledge graph as context routing. Part 2.
Why AI needs architecture, not just a powerful model: memory, knowledge graph, retrieval, bounded delegation, and verification for reliable AI assistants. Part 1.