Rules for matching bank lines and invoices can't learn, so I kept writing new ones. Now code builds a short list, a decision model picks from it, and a Cloudflare Workflow waits for a person when the model isn't sure.
Our customers ask Claude and ChatGPT about their solar plants. Cloudflare Access and an MCP server portal handle who gets in, so we didn't have to build auth. The cf CLI lets my agent set it up without a browser.
Wawa used to send every saved WhatsApp batch to the task model. Now a single decision call picks run, wait, or skip first, and anything it isn't sure about still runs.
Wawa's Spotlight searches saved messages, images, audio and video with EmbeddingGemma 2 on a local Ollama. No GPU, no transcripts, and the index lives in the workspace database.
I measured a Tauri and an Electron window for Wawa's desktop app. The window was smaller. The whole app wasn't, until the sidecar moved from Node to Bun.
Wawa used to store every item as a task. Now it keeps what an item is, where it stands and when it matters as separate fields on one record, with waits stored apart.
Our Dagster steps kept fighting over the same Postgres rows. Now they append events to Cloudflare Basin streams, a sink lands them in Iceberg, and the reader decides which version wins.
Users write their own threshold rules. Each site gets one Durable Object that checks them. The database owns the rules, the object owns the checking, and a version number keeps them honest.