"Google Cloud offers a very high degree of flexibility, such as selecting the number of GPUs from 1 all the way up to 8. Previous cloud providers didn’t offer this level of granularity — forcing us to start with a fixed number of GPUs," explains Tan. Being restricted to a few types of instances led to wastage, as some stages of code development didn’t fully use the GPUs that came with the fixed instances. With Google, we can start small and scale up as required.

Google Cloud’s reliability has been a determining factor for the team — when it comes to training Large Language Models (LLMs), any interruption could wipe out months of work and resources. Google Cloud Storage and Persistent Disk provide the scalability and performance needed to handle the vast amounts of data involved in LLM training, from preparing datasets to managing model outputs. "The data management tools from Google Cloud have been a real turning point for us. They’ve made moving models and data between instances so much easier, cutting down on time that would’ve been wasted on reinstallation and transfers. This has improved our workflow and helped us keep our projects on track," says Tan.

Pints.ai containerizes its machine learning models and runs them on Google Kubernetes Engine (GKE) to ensure consistent environments across development and production. Thanks to GKE’s scalability, Pints.ai can handle fluctuating demands without compromising on performance, while reducing the operational burden of managing infrastructure. 

By integrating Vertex AI into its end-to-end machine learning pipeline, the company automates essential steps in model development, such as hyperparameter tuning — adjusting model parameters to improve performance — and model evaluation, which ensures the model’s accuracy and effectiveness before deployment. "Vertex AI has been instrumental in helping us deliver faster, more accurate, and tailored solutions that meet the unique needs of financial institutions. We also incorporate Retrieval-Augmented Generation (RAG) to allow financial professionals to query vast datasets and document repositories using natural language," says Tan.

Using Gemma Open Models alongside Google Cloud’s advanced GPUs, Tan’s team can train models faster and with better results. Gemma pre-trained models provide a head start for developing custom AI solutions, so they can focus on fine-tuning models specifically for financial use cases. Pints.ai made their open-source model, one and a half pints, publicly available as part of a research initiative. They provided all the necessary code, along with detailed instructions, so that users could easily train the model on Google Cloud in under 10 minutes. "The idea is not so much about replacing Gemma or Vertex AI, but offering an option for cases where general open-source models don’t satisfy specific needs. This is where our small model comes in, and you can do it too — it’s very easy to do on Google Cloud," says Tan.

"Our collaboration with Google has been a key factor in building our generative AI capabilities. This partnership has provided incredible support and access to computational resources, which has become a pillar of our AI initiatives," says Tan. Moving forward, Pints.ai aims to deepen the collaboration with Google Cloud to further scale their AI operations. This includes increasing utilization as well as scaling up to more powerful instances provided by Google Cloud.

They have also made a commitment to actively share their experiences and insights with the broader community, starting with how they’ve used small model templates to empower other companies in building and optimizing their AI models. "Our goal is to contribute to the development of best practices in AI and cloud adoption, drawing on our successful implementation of Google Cloud," says Tan.