AfriRamp AI: A Fine-Tuned Gemma Assistant for African Freelancers

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle.

What I Built

AfriRamp AI is a web-based AI assistant designed to help African freelancers and remote workers understand cross-border payments and the AfriRamp financial service concept.
It inspired to help users pf Afriramp product

Freelancers working with international clients often face confusing payment options, high transaction costs, delays, and difficulties converting foreign earnings into usable local currency. I wanted to explore how a small, customizable AI model could make financial-service information more accessible to this audience.

Instead of relying on a closed, general-purpose AI API, I fine-tuned Google's Gemma model to build a domain-focused assistant that can answer questions about AfriRamp and its intended users.

The project is an early working prototype. It currently provides conversational guidance about the AfriRamp concept; it does not execute payments, exchange currencies, or move money.

The potential business model is to offer a useful financial-information assistant to freelancers, with a future opportunity to provide premium features, localized guidance, and integrations with payment services as the underlying product develops.

Demo

Try AfriRamp AI: https://afriramp-ai-assistant.vercel.app/

The frontend is deployed on Vercel and connected to a FastAPI backend serving my fine-tuned Gemma model.

Ask it questions such as:

• What is AfriRamp?
• Who is AfriRamp designed for?
• How could AfriRamp help freelancers receiving international payments?

The model backend currently runs in a Kaggle notebook and is exposed through a Cloudflare Quick Tunnel. This is a prototype deployment, so the demo depends on the notebook and tunnel remaining active.
The tunnel end point is: https://gaps-charity-absolute-peoples.trycloudflare.com

Code

GitHub repository: https://github.com/kaksv/afriramp-ai-assistant

The project brings together a web chat interface, a Python API, and a fine-tuned open-weight language model.

How I Built It

The core of this project is Google Gemma 3 270M IT, an instruction-tuned, open-weight language model.
The file used for training the model can be accessed here

I fine-tuned the model in a Kaggle notebook using a Tesla T4 GPU. For this initial experiment, I prepared 40 synthetic training examples and kept another 8 examples aside for evaluation.

I then built the application around the fine-tuned model:

• Model: Google Gemma 3 270M IT, fine-tuned for AfriRamp-related conversations.
• Training and inference: Python, PyTorch, and Hugging Face Transformers.
• Backend: FastAPI exposes a /chat endpoint that receives a user's question and returns the model's response.
• Frontend: React and Vite provide the browser-based chat experience.
• Deployment: Vercel hosts the frontend, while Kaggle runs the model and Cloudflare Tunnel exposes the API over HTTPS.
• Development: GitHub and GitHub Copilot supported source-code management and frontend implementation.

Getting the pieces to work together was an important part of the project. I tested model inference first, exposed it through FastAPI, verified the endpoint, and then connected the deployed frontend to the live API.

The current model and dataset are an initial experiment, not a production-grade financial assistant. Further evaluation with real user questions, better training data, and more rigorous safeguards will be needed.

Why Does Open Innovation Matter?

Open innovation made it possible for me to experiment with a domain-specific AI product without building a large model from scratch or depending entirely on a paid, closed-model API.

Gemma gave me a model I could inspect, adapt through fine-tuning, and integrate into my own application. That flexibility matters when building for a specific audience whose needs may not be well represented by generic assistant behavior.

It also let me learn across the complete AI development process: preparing training examples, fine-tuning a model, testing inference, building an API, and deploying a working interface.

For a solo builder working with limited resources, this approach lowers the barrier to experimentation. It gives me greater control over the model and the opportunity to improve it as I learn from users.

The project is not yet fully self-hosted or production-ready, and it still relies on hosted infrastructure. However, the central AI component is a model I can customize rather than simply consuming a closed API.

My long-term goal is to explore how accessible, adaptable AI can support African freelancers and remote workers navigating cross-border financial services.

Prize Categories

• Best Use of Gemma
• Best Use of GitHub Copilot