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 - and don't just take my word for it: all of them are live on our demo page, where you can switch between them as often as you like.
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).
🍳 Kochmonster (kochmonster.de):
The resident chef of a German recipe and cooking site with a huge vault of recipes - some of them straight from star kitchens. Opinionated, a little full of himself, and absolutely devoted to serious cooking. He will happily tell you what's wrong with your sauce, and he does it in German.
What makes him more than a fun character: he doesn't invent recipes. Through an MCP plugin he searches the site's live recipe vault, editorial articles and cooking knowledge base - filtered by time, course, cuisine and difficulty (which runs from Apprenti to Chef de Cuisine, naturally) - and sends you straight to the recipe page.
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).
For the actual chat response generation, ConverZen now talks to OpenAI, Anthropic's Claude, Google's Gemini and Groq natively - plus any provider that implements the OpenAI interface, from DeepSeek to Venice. Groq takes the response times of a chat down to blistering fast, below 2-second waits: the bot streams text at 480 tokens per second. If you have a need for speed, this is who you want to talk to. 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. Kochmonster shows what that looks like in practice: ask him for a quick vegetarian main course and he pulls matching recipes from the live site instead of making one up. 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.
It started with OpenAI. Then came Groq's LPUs, Anthropic's Claude and Google's Gemini - and with them every provider that speaks the OpenAI interface, like Venice or DeepSeek. Which one handles Zeno's dry nerd sarcasm best? That's now a per-persona setting, not a rewrite.
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 - or first go and meet the bots yourself. Each one keeps its own conversation, so switch as often as you like.