Blog

I shipped a couples app in eight days. Eight people signed up, the last on 14 July, and nobody has answered a daily drop since the 16th.

Ragnarok: building a multi-tenant gym operating system from first commit to gym-wide production in seven weeks, and migrating 441 members off the incumbent.

Autonomy isn't a property of the agent. It's a budget you spend per action, priced by how reversible the action is and how far the damage could spread.

We built a no-code agent builder so SMEs could build their own agents. Almost nobody did. Why Studio failed, and what we built instead.

Templates handle the common 80%. The remaining 20% is every customer's actual business. Here's what that 20% looked like for one bakery, one day, five bugs.

Three lessons from selling AI to businesses at Voltade: you're selling trust, be yourself, and notice when the customer starts selling to you.

Model-layer safety is the lab's job. Deployment-layer safety is mine: pointing a safe model at a real customer's WhatsApp without getting burned.

What an AI interaction actually is, which layer a task-success rate is measured at, and the eval, guardrail and cost checks that keep both numbers honest.

Voltade's third swing at the same problem: Studio asked SMEs to build, Vobase had us build for them, Volty bets they'll configure. What changed each time.

Agent safety in production isn't about jailbreaks. It's an agent confidently doing the wrong thing on a customer's WhatsApp for hours.

Seven personal agents on a Mac Mini, four months of breakage, and the two fixes that stuck: scope discipline and deterministic flows.

Most agent evals measure the wrong things. The framework I use on two production agents, LLM-as-judge calibration, and the $300 mistake that started it.

A crypto trading bot built with Claude Code, $100 in USDC and a lot of guardrails. It lost everything, so I rebuilt the strategy from actual research.