Ingest pipeline, semantic embeddings across hundreds of thousands of records, cloud deployment on GCP and Neon — in about two weeks using Claude Code CLI. No IDE. Then I had to deliver it to a non-developer colleague who uses Claude Desktop on Windows. This essay covers the sharp edges of that transition: environment variables that don't propagate from dock-launched apps, PgBouncer silently dropping data during batch jobs, Windows path escaping in JSON config files, and the pragmatic reason MCP exists (it's the only mechanism by which Claude Desktop's Chat tab can call external tools — without it, there's no data access). Happy to answer questions about the architecture or the delivery process.
A fake academic paper fed to NotebookLM's Audio Overview to see what happens when two AI podcast hosts discover that all male-female co-hosts inevitably fall for each other — and are asked to consider what that means for them. Seven fake studies, 4,218 fake dyads, very real awkwardness.
A Substack post that includes four preview chapters from the first full English translation of Zhang Henshui’s The Story of a Noble Family, now out as an ebook on Kobo. The entire project was built and run locally on a Mac Studio (M2 Ultra, 192 GB RAM) using Python scripts with Ollama-served Llama 3 70B to segment and translate the original 1920s Chinese text. The pipeline consists of regex chapter parsing, style-aware prompting, and offline batch translation.