One long-lived Claude Code session per project, woken by dropping a request file in a folder. Each project keeps one pinned conversation that gets revived after a crash or reboot, and a chief-of-staff agent hands issues to it within a cap on busy sessions.
A remote MCP server that lets Claude's phone app, voice mode included, see and act on the lab: what's waiting on me, mark a task done, hand an issue to a coding session. Every change is read back and waits for a yes, retries never apply twice, and nothing at home has an open port.
An agent that runs my notes and tasks. I capture and decide; it triages the inbox, turns plain sentences into task changes, writes a morning brief and mirrors GitLab issues into an Obsidian vault. It started with a migration of ~1,900 Evernote notes, checked attachment by attachment.
A self-written live-trading layer for the open-source Jesse framework: an exchange order driver, a real-time WebSocket candle feed and session reconciliation, with paper trading on the same code path. Building it turned up and fixed real engine bugs. About the engineering only, never returns.
A domain assistant built on LangGraph: a classifier routes each question to SQL analytics, web research or a vector knowledge base. It lives inside Open WebUI or as an embeddable widget with per-app keys, and GitLab issues can be handed to headless Claude Code runs that open merge requests.
One publisher for my content channels: it keeps each channel's posting calendar, schedules reels to social platforms, levels the audio and remembers what already went out.
Tried something from the lab, spotted a mistake, or just want to talk automation? I'd like to hear it.
[email protected] →Paid work goes through my company, so you get proper terms and invoices.
[email protected] → via Shiloah Software SolutionsCan't decide? Either works, I read both 🙂