Blog

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.

The two numbers on my CV, unpacked: what counts as an interaction, which layer the 99.65% is measured at, and the eval, guardrail, and cost subsystems that keep both honest.

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

Production agent safety is not about jailbreaks. It is about an agent confidently doing the wrong thing on a customer's WhatsApp for six hours before anyone notices. Here are the failure modes that actually happen, the guardrails that work, and how I prove the guardrails are working.

Most agent evals measure the wrong things. After running two agents in production for six months, here's the framework I actually use, with real metrics, LLM-as-judge calibration data, and the $300 lesson that started it all.

Autonomy isn't a property of the agent. It's a budget you allocate per action, priced by how reversible the action is and how big the blast radius gets. Here's the framework I use across every agent I run.

An agent that can rewrite its own behaviour can rewrite it wrongly. This is the machinery that decides which of its proposed changes land on their own, and which have to wait for a human.

An 8am cron reads 24 hours of my AI agents' logs through four lenses and DMs me a prioritised list of what to fix. Plus the open-source self-healing harness underneath it.

Volty agents draft, humans approve. WhatsApp gives you 24 hours to reply. What happens when the approval is slower than the window, and the keep-alive sweep we built for the collision.

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.

The mechanic I actually use to prioritise across two 0-to-1 AI products as the only PM at Voltade. The Studio pivot, the Thursday rule, and the call I got wrong.

Eighteen months building applied AI for SEA SMEs at Voltade. Here's what I'd build first if I were starting Singapore's new applied AI lab.

Nous Research open-sourced Hermes Agent, a harness in the same shape as OpenClaw. I read it for an afternoon and ported four patterns: a usage tracker, trajectory compression, a stdio MCP server, and an error classifier in the claude-cli shim. Two more I deferred. Two I rejected outright.

Garry Tan's gbrain has a few design moves that solve real OpenClaw pain. I ported five of them: a durable job queue, a config doctor, storage tiering, a hot-memory recall CLI, and a compiled-truth memory structure. Here's what worked, what didn't, and what almost leaked to a customer channel.

A walk through the self-learning loop inside Vobase: wake events, staff signals, change proposals, applied skills. The technical mechanism that turns adaptive software from a slide into a thing SME owners actually trust.

I run seven personal AI agents on a Mac Mini in my flat. They book my gym classes, manage my inbox, handle outreach for two academies, and watch Voltade customer groups. The harness took four months to stabilise. The two things that fixed it were scope discipline and deterministic flows. Here is what each agent does, what broke, and the patterns that finally stuck.

I did a full stability audit of my OpenClaw stack today. Here's what changed once it stopped being a personal toy and started serving real customers and a real team.

Notes on writing a model behaviour spec for an agent that works for someone. What it does, what it refuses, who it serves when those conflict, and why most system prompts are a lazy substitute for a job description.

Notes on when to use Haiku, when Opus is worth it, and why model choice is a product decision wearing engineering clothes.

82 sessions, 12 weeks, 4,888 Bash calls. I pointed an audit tool at my own Claude Code transcripts. Three things came back that I didn't want to hear.

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.

Fixed SaaS loses to models. Blank-canvas AI builders lose to the cold start problem. What wins is 80% ready software that the business shapes by talking to it.

Every time you open Claude Code on a new repo, you start from zero. claude-init fixes that by analysing your codebase and generating a complete .claude/ configuration.

I built a Claude Code skill that writes blog posts in my voice from any session. The hard part wasn't the workflow, it was the voice.

How we built Envoy from zero to 20+ paying SME customers in Singapore. Dogfooding, grant-driven GTM, killing features, and the product decisions behind it all.

This entire website was built, redesigned, and maintained by Claude Code. Here's every prompt, tool, and workflow behind it.

I built an agent observability dashboard after an OpenClaw cron job silently burned $300. Here's the problem and what WIMAUT does.

How I set up a local LLM on a base-model Mac Mini, put it behind a public URL with Cloudflare Tunnels, and locked it down with Cloudflare Access. No port forwarding required.

I built two OpenClaw agents for Voltade. One manages our customers. The other runs our internal ops. Here's what I learned about deployment, models, and what agents are actually good for.

Software's marginal cost is approaching zero. SaaS as a business model is dying. The winners will be AI-native services firms, not tool makers.