I build products the way directors make films.

New product, not a feature. A separate launch with a migration path.

I explored the shoehorn path and did the design work. It didn't hold up. The existing product's philosophy — simplicity through constraint — is incompatible with an AI agent's unlimited-behavior model. A demo settled the stakeholder conversation.

If the product is AI-native, the configuration should be too.

The design decision followed directly: asking SME owners to fill in forms and toggles to configure an AI agent is a category error. The primary interface should be a conversation — not a settings panel with an AI helper bolted on.

Structured views as read-back surfaces, not the primary entry point.

The chat interface drives configuration. But SME owners sometimes want to verify what they've set up. Routing rules and the directory persist as structured read-back views — the conversation generates the data, the panels expose it.

The agent is the brand's voice — literally.

A law firm sounds different from a surf shop. The persona picker lets SME owners choose and audition a named voice that matches their business, making the agent feel like a capable colleague instead of a generic configuration panel.

Built a self-improving feedback loop — not just a product.

Every call session is an implicit test. An automated eval layer runs deterministic checks and AI-judged quality assessment on every interaction. Findings surface as actionable cards. The next step: route those cards to coding agents for automatic remediation.