Why do 99% of corporate AI chatbots sound like exhausted customer service reps on valium?
"How can I assist you today?" – If I read this default prompt one more time, I'm going to fall asleep.
Brand voice matters. A local windsurfing shop shouldn't sound like a corporate bank in its chat UI. That's why I stopped building standard AI bots and started building characters.
Meet the extremes of my current production environment:
đź”® Serena (gaylynmorgan.com):
An esoteric, highly empathetic Tarot expert who will patiently discuss life, love, sorrow & joy with you. She gives counsel on strategies to overcome whatever is holding you down—all in the context of Gaylyn's philosophy. Even if it's just recommending one of her services and leading you to the product page. Talk to her if you need cheering up.
đź’» Zeno (converzen.de):
My Lead Engineer & Nerd-Bot. A cynical, impatient Senior Rust Developer who hates small talk, demands efficiency, and will roast you if your latency is too high. He might not be the perfect fit for a standard sales audience, but I crack up laughing every time I talk to him—and he knows all the facts. (I will probably have to put a slick sales pro at his side soon).
The secret here is to supply so much actual context that you keep the LLM from hallucinating and making things up that your business cannot deliver. You have your sales department for that.
But a fun personality doesn't drive business on its own. To keep this at the absolute bleeding edge, the underlying architecture has to be flawless:
⚡ Zero Lag:
ConverZen's backend is written entirely in Rust—not a single line of TypeScript. Memory-safe, insanely fast, and enterprise-grade secure. Everything other than the web UIs.
We just committed a backend service to production that uses locally run models (in candle/rust, of course) to do moderation and embeddings for the chat and RAG system. This task is now typically finished in 100 ms instead of the 4-5 seconds it took when external models were handling it. (If you need faster, we can wire in a couple of GPUs).
Currently, we are implementing my second LLM engine interface (OpenAI is the first) for the actual chat response generation. Groq takes the response times of a chat down to blistering fast, below 2-second waits. There is still some work to be done to keep my chat personas in character across models & providers, but if you have a need for speed, this is who you want to talk to. With Groq, the bot streams text at 480 tokens per second. The loading spinner is dead.
Custom RAG:
Retrieval-Augmented Generation injects deep domain knowledge into the chat. Your philosophy, your messages, your vision, your product descriptions, manuals, FAQs, and Readmes can and should all be part of your bot's knowledge base. RAG injects morsels of information into the chat that are relevant to the user's message, telling the LLM what to concentrate on instead of letting it invent its own opinions. That way, it can leisurely chat about your products, why it is cool to purchase from you, and what sets you apart from the crowd.
MCP (Model Context Protocol):
This is the ultimate e-commerce weapon. The AI doesn't just talk about your realm and your products; it queries your actual shop database in real-time to pull prices, stock, and direct product links. It knows your webpage inside and out. "Where can I get support for this product?" If the bot doesn't know the answer itself, it can point your customer to the exact link. Currently, in ConverZen this is still mostly old-school (but blazingly fast) fuzzy keyword searches. With our local models now in place, we will soon have intelligent semantic searches while trading 20ms MCP replies for 50-100ms latency.
The Bleeding Edge
In 2026, new and better models evolve faster than I can type
#[async_trait]
impl RGHandler for OpenAIResponses {
async fn generate_response(
&self,
messages: Vec<SessionChatMessage>,
) -> Result<RGResponse, RGError> {
todo!()
}
That is exactly why ConverZen is built from the ground up to be completely model-agnostic.
Being locked into a single AI provider right now is an architectural death sentence. Because my Rust backend handles all the heavy lifting locally—the session memory if required, the custom RAG vector searches, the MCP tool execution, and the toxicity moderation—swapping out the actual "brain" of the chatbot is literally just changing an API endpoint and a JSON payload.
Yesterday it was OpenAI. Today it's Groq's LPUs. Next up on my hit list are Anthropic's Claude and Google's Gemini, just to see which one handles Zeno's dry nerd sarcasm the best.
The AI landscape isn't slowing down, and neither am I. If you're an e-commerce techie looking to future-proof a shop and take it from "Hey..." to "Wow!", let's talk.