Assistant Sona

The problem

AI tooling is fragmented. Getting through an ordinary set of tasks means moving between several services, each with its own interface, its own quirks, and its own learning curve. The tools are good. The switching cost is what wears you down.

I wanted to find out whether a single, focused interface could cover the most common productivity cases well enough that the switching stopped being necessary.

What I built

Assistant Sona brings five AI-powered tools into one interface with a shared interaction pattern, so you learn the interface once instead of five times.

• Tool one — what it does
• Tool two — what it does
• Tool three — what it does
• Tool four — what it does
• Tool five — what it does

The decision that mattered

My first instinct was to route each tool to a different model depending on the task. It sounded sophisticated. In practice it made output quality unpredictable and the architecture harder to reason about.

I swapped model routing for tightly scoped prompts. Simpler system, more consistent output, and I could actually explain why it behaved the way it did.

That trade — less cleverness for more predictability — is one I'd make again.

What I'd do differently

• Handle API rate limits and timeouts properly rather than letting them surface as failures
• Write tests for the prompt-to-response pipeline instead of testing by hand
• Cache repeated requests to cut both latency and cost