Why a language model isn't the same as an AI assistant, what an app contributes, and how a clear assignment and reliable information turn technology into something useful.
A question worth testing
As RAB Products added products and educational experiences, visitors needed an easier way to find answers. But we didn't want a chatbot that interrupted everyone. Could we create a small assistant that helped only when invited?
On October 27, we started experimenting with PocketPal AI on an iPad and a compact model from the Qwen family. No expensive laboratory, and no attempt to train a new model from scratch.
Three pieces of the experiment
The model was the language engine. Trained beforehand on large amounts of data, it could generate responses—but did not automatically know our business, our products, or whether a particular answer was current.
The app let us run the model and converse with it directly on the iPad. PocketPal provided the local interface; it wasn't itself the model, and an internet connection wasn't required for those local conversations.
Our instructions and information supplied the intended job: answer courteously, represent RAB Products accurately, and admit uncertainty. This is a crucial distinction. Writing instructions does not retrain a model. It guides how the existing model is used.
Our first tests exposed limits, too: the iPad struggled with demanding models, and the assistant could confuse who was speaking. These weren't reasons to abandon the experiment. They showed us what needed testing before the system could represent a business.
We gave him a name
On October 28, 2026, the experiment became Chipper. The name didn't give the model consciousness, feelings, or new facts. It gave us an identity for an assistant we could deliberately design and evaluate.
We chose a friendly blue-and-silver shield face, a thumbs-up, and a reserved place in the website corner. More importantly, we gave Chipper boundaries: be approachable, avoid unsupported claims, and say “I don't know” when appropriate.
From a local experiment to a website assistant
By October 30, Chipper had his first assignment on the RAB Products website. The visitor-facing service used Cloudflare infrastructure rather than simply exposing our original iPad experiment to the public. The two environments served different purposes: one helped us explore; the other made assistance available to visitors.
We also investigated web-search tools such as Brave Search and You.com. Considering a service isn't the same as adding it to the live assistant. Each tool needs a purpose, safeguards, and testing.
Building an AI assistant doesn't require inventing a new language model. It means choosing an existing model, giving it a useful role, surrounding it with dependable information, and checking whether it actually helps.
A beginning, not a finish line
Chipper now had a name, a face, and somewhere to work. Could he keep his information current, recognize an error, use tools safely, and improve through visitor feedback? Those questions lead directly into the next chapter.