This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
I chose this project because as a kid (since age: 17) I have always had this vivid dreams of speaking with Elon Musk, he is kind of like my Idol and I would cherish the opportunity to have a conversation with him, until that day comes, this is what we would manage. When I came across this hackathon to test the lates model built by Google (Gemma 4) it was a no-brainer for me. I want the opportunity to understand the memory and context management of this model. I decided not to limit it to Elon Musk and included as much celebrities as possible. I want to give my friends the opportunity to test and use it.
Buddy is a web application that lets you have real, in-character conversations with iconic public figures such as athletes, billionaires, historical leaders, artists, and innovators. You pick a person, and the AI steps into their shoes completely. It speaks the way they speak, stays grounded in their real life and career, and refuses to break character.
Although, with this project I am not solving a problem like climate change or malaria cure, The problem Buddy solves is simple. Most AI chatbots are general-purpose. You can ask them anything, and they will answer as a neutral assistant. That is useful, but it is not immersive. Buddy flips that entirely. The AI is not an assistant anymore, it is a person. Ask Lionel Messi about his childhood in Rosario. Ask Steve Jobs why he believed design was everything. You get answers that feel like a real conversation, not a Wikipedia summary read back to you.
The experience Buddy creates is curiosity without friction. It is the feeling of sitting across from someone remarkable and just... asking them things.
Demo
Link: https://ifeanyi-buddy.netlify.app
Code
GitHub: https://github.com/MasterIfeanyi/google-gemma4-hackathon
How I Used Gemma 4
Buddy runs on Gemma 4 26B, accessed via OpenRouter using the google/gemma-4-26b-a4b-it:free model.
The choice of Gemma 4 over the smaller variants came down to one core requirement; persona consistency. Keeping an AI locked inside a specific person's personality across a full multi-turn conversation is genuinely hard. The model needs to remember who it is, match the person's tone and vocabulary, deflect off-topic questions naturally without sounding robotic, and stay grounded in real publicly known facts all at the same time. The smaller E2B and E4B variants are impressive for their size, but they tend to drift out of character more easily under pressure, especially when a user deliberately tries to break the persona with trick questions or unrelated topics.
The 26B model holds the character much more firmly. When you ask Messi a question in Spanish, it does not just translate, it responds with the rhythm and warmth of how Messi actually speaks. When you push Steve Jobs on a topic he would find irrelevant, it redirects with the kind of sharp confidence Jobs was known for. That level of nuance needs a bigger model.
One specific technical challenge worth mentioning is Gemma 4 does not support a native system role in its message format. To get around this, Buddy injects the personality prompt as a user / assistant turn pair at the very start of every request, effectively tricking the model into treating the persona as established context before the real conversation begins. The 26B model responds to this pattern reliably and stays locked in. Smaller models were noticeably less consistent with the same approach.
What Gemma 4 ultimately unlocked for Buddy was believability. Because the model is large enough to hold complex persona context across many turns, we never had to simplify the experience to make it work. We did not have to limit conversation length, restrict the types of questions users could ask, or hardcode fallback responses. The model handled edge cases we never explicitly planned for. When a user asked Messi to speak Spanish, Gemma 4 switched languages naturally and stayed in character at the same time, responding with the warmth and rhythm of how Messi actually communicates, without any instruction from us to do so. When users tried to break the persona with off-topic questions, the model deflected in the voice of the person, not with a generic AI refusal. That level of emergent, in-character behaviour is what a smaller model cannot reliably give you and it is what makes Buddy feel like a conversation rather than a demo.