Engineering
Notes from building Content Shifted: architecture decisions, agent loops, and how features actually ship.
Implementation sketch: chunking, embeddings, semantic search, metadata filters, and reference links on top of the existing knowledge edge.
Read article →What a DAG is, how LangGraph looks in pseudocode, and how my canvas is a durable automation graph (media today; same shape for Sentry triage, GitHub tickets, and other flows), with loop engineering inside nodes and multi-agent wiring via edges.
Read article →Tradeoffs between an LLM librarian and a vector index for knowledge context, and why I bet on vectors (with metadata filters), including when an agent still needs to pick a file.
Read article →How we turned a messy multi-step image pipeline into a focused agent: three tools, finish-time QA staging, and retries that improve overlay readability instead of guessing forever.
Read article →Shared inputs plus a uniform output (always a file, plus step/critique logs when agentic) so canvas actions, an in-app agent, and a future MCP server stay thin wraps, not rewrites.
Read article →Workflow presets materialize a full automation graph from a declarative config, and website presets bootstrap logo and brand styles from the live site in the background.
Read article →