Turning an Obsidian vault into a local AI engineer: redact docs before agents see them

AI Engineer · youtube · 2026-10-11

Filip Makraduli (Superlinked) demos a workflow built on an Obsidian vault, MCP and Superlinked's inference engine: a scanned NDA gets OCR'd, redacted and converted to a small Markdown artifact before any coding agent touches it, avoiding privacy leaks and context costs of feeding raw pages to large models. The stack splits into host, MCP server and GPU worker pools handling model loading, batching and shared workloads, with embeddings, reranking and task-specific LoRAs. A live section stress-tests the privacy claim — if the agent can read raw files first, instructions alone may not hold; fixes include separating file access from processing and running a private cluster.

Original post →

More from coding & agent

coding & agent channel →