How To Build an AI Chatbot
We sat down with VienerX Generalist, Bill Lioreisis and CEO Wayne Viener to discuss the process of creating an AI chatbot. Wayne has been working on bringing a “CPA character” named Hal to life using ChatGPT tools.
Hal Brenner is a 68-year-old accountant who was born in the Philadelphia area. When he was 10, his family moved to Miami Beach, and as a teenager, they relocated to Manhattan. After attending college at Georgetown University in D.C., he settled in Northern Virginia, where he married his wife, Ella, and they raised two kids. Today, Hal is retired and living the good life hitting golf balls in Boca Raton, Florida.
If you ask him anything about his life, he’ll be happy to tell you. About how he grew up an New York Giants fan, but after graduating from Georgetown in 1979 and living in the D.C. area for years, he eventually switched over to being a Commanders fan. Although he’ll fully admit, when the two teams play, he is torn.
About his kids, Rachel, a lawyer in New York, and David, working in tech finance in Silicon Valley, who are the pride and joy of his life. If you spend some time talking to Hal, it’s not a stretch to feel like he’s a real, witty, Grandpa who is enjoying the fruits of retirement.
Except Hal isn’t real. He’s an AI persona built using ChatGPT (technically, a GPT). VienerX’s Wayne Viener spent months working on him, and was later joined on the project by Bill Lioreisis. This is the story of how Hal came to life.
Beginnings
This project started very differently than its current form. Wayne, who has spent much time learning about the cutting edge of AI, decided it would be an interesting challenge to develop an AI bot to help with financial questions. Wayne didn’t just want the bot to be a boring financial GPT though, he wanted to give it personality, and he wanted him to have persistent memory, the technical term for the ability to remember what you tell him last time. That eventually became Hal.
From there it developed further. Over the course of the next few months, Wayne, who was joined by Bill, began to conceptualize the idea as something that could be marketed to customers.
The Story Side
The story, the actual personality, was Wayne’s doing. He compares it heavily to story writing, something he has a lot of experience in through his second career as a sports journalist. Those who know him can see the personal touches in Hal. The Philadelphia in his background mirrors Wayne’s older son, Jordan’s years living in the city. While Hal’s current home in Boca is where Wayne’s mother has lived off and on since 1982.
When writing the details for Hal’s story, one thing that Wayne was very careful about was pouring enough detail into Hal that he would be perceived as a real person. He estimates that there are several thousand words put into creating the background of Hal’s personality. That said, even with putting all the effort in, one thing the team emphasized was stress testing it.
Most AI tools are based on Large Language Models (LLMs). LLM AI has a tendency to do what is called, “hallucinating.” Not in the human sense, although similar. Hallucinating in this sense, is having the GPT generate a piece of false information and presenting it as factual. This can be because of technical failures or other reasons. These hallucinations can be difficult to discern from the truth, particularly if you as the user don’t already know what the answer is. For this reason, as ChatGPT itself always says at the bottom of the screen, it is always critical to verify important information. If you are suspicious of the information that your doctor gave you, you seek a second opinion. People must be willing to take that same level of skepticism, perhaps more, with their AI agents.
Another common thing that had to be worked out with testing was consistency of answers. A simple question like “where do you live?” can easily be inconsistently answered, particularly when Hal has “lived” in five locations throughout his life. Like any other AI source, it is important to always double check information when using even a specialized bot.
The Technical Side
ChatGPT, the tool behind Hal, allows developers to access an Application Programming Interface (API). Think of the API as the “vanilla” version of ChatGPT. When a user interacts with ChatGPT using their account, they’re effectively customizing it with their own data. This same philosophy applies when designing customer AI bots, albeit, with more steps.
For Hal, the goal was to have the AI remember specific details and use them in every interaction, this is known as persistent memory. Achieving that required a process called “wrapping”, building a structure for the GPT’s memory to live in. This was no small task. The logic of AI memory is complex. For example, the developers had to program it so that after each interaction, Hal could access relevant portions of memory and summarize them accurately in about 100 words. This foundation is what allows Hal to function as a consistent and believable persona.
The Future of Hal
Hal is close to being ready for prime time. The team at VienerX learned a lot throughout his development and will soon be ready to build AI bots for customers. When asked how this story should end, Hal itself responded, “Technology may do the math, but it’s human judgment that gives it meaning. I’m here to keep that judgment alive — just faster, clearer, and maybe with a little more Boca charm.”
