The durable-state plumbing that makes human-in-the-loop actually safe in production.
|
Diagnosis before solution. Most "performance issues" turn out to be visibility issues; most "feature requests" turn out to be a missing constraint. The fix is worthless if it's aimed at the wrong bottleneck. Multi-step pipelines with explicit lanes, durable state, layered guards, and human-in-the-loop where it matters. Account safety beats growth, every time. Small teams, ambiguous problems, real production constraints. The work that lands is rarely the work that was specified — so the system has to be cheap to redirect, not expensive to defend. |
|
OpenClaw is my own local-first agent runtime — loopback gateway, persistent Chrome session, fleet of specialized agents behind a layered security guard.
Built and shipped daily through the AI tools in the row up top — Claude Code and Codex carry most of the weight.
|
Most of the work that actually stretches me — vault crypto, agent security, an EHR surface, AI pipelines for media ops — lives in private repos. The public surface here is intentionally lighter than the day-to-day reality while I build the systems that will write the proof out on autopilot. So a few things worth knowing:
|
→ Most active on X · @shahid_bagwann
daily engineering notes · teardowns · what I'm thinking through, before it lands anywhere else