OpenPipe

OpenPipe is a Seattle-based company that operates within the B2B industry, specifically focusing on engineering, product, and design. With a small team of two, OpenPipe participated in Y Combinator’s S23 batch and actively serves regions across the United States and Canada. The company’s core innovation lies in abstracting the fine-tuning of custom models by capturing existing providers’ prompt-completion pairs to generate new models. These new models are not only faster and cheaper but often surpass the accuracy of the original models. OpenPipe offers a software development kit (SDK) that supports Python and NodeJS, facilitating easy integration and usage for developers. The SDK allows for the pruning of static text, which reduces input token counts and thus saves on inference costs. OpenPipe’s system supports tagging for better data management and automatically records LLM requests and responses to aid in training. Notably, the fine-tuned models inherit all pruning rules applied to their training dataset, ensuring that any rules added post-training do not affect existing models, thereby maintaining backward compatibility. OpenPipe also encourages open-source contributions through its GitHub repository.