Hey there! I’m Kyle, the founder of OpenPipe. OpenPipe is the fully-managed fine-tuning platform for developers. Our users have already saved over $7M while lowering latency and improving quality by switching from GPT-4 to fine-tuned models on our platform.

Today I’m excited to announce the close of our $6.7M seed round. The round was led by Costanoa Ventures with participation from Y Combinator, Logan Kilpatrick (former head of DevRel at OpenAI), Alex Graveley (creator of Github Copilot), Tom Preston-Werner (founder of Github), Flo Crivello (founder of Lindy.ai), Immad Akhund (founder of Mercury), Goodwater, and many others. We are grateful for their vote of confidence!

What is OpenPipe, and why should I care?

OpenPipe is the easiest way to replace your GPT-4 prompt with your own fine-tuned model. Fine-tuned models are cheaper, faster and more reliably correct compared to a crafted prompt. For projects and businesses that make at least 1000 calls to an OpenAI prompt a day, we’ve found that switching to a fine-tuned model is a true Pareto win—you can improve on all 3 axes of speed/cost/quality, sacrificing nothing.

OpenPipe makes it extremely easy to get started. All you need is an existing prompt in production — use OpenPipe to collect your production data, filter and relabel, train a model, and deploy it with an OpenAI-compatible endpoint. We’ve seen users with no ML experience successfully deploy their own fine-tuned models in less than an hour.

OpenPipe’s usefulness grows even more once you’re running your fine-tuned model in production. That’s because we’ve built the first platform to automate your data flywheel, helping you create a model that becomes stronger as your user-base grows.

The new data flywheel

Beyond the initial wins of deploying a cheaper and faster fine-tuned model, OpenPipe helps you build your data flywheel. A data flywheel is a technique large enterprises with sophisticated data science teams use to widen and deepen their moat. For example, Amazon is able to provide better product recommendations than an upstart competitor because it has valuable data about how often existing customers click on each product after seeing it in a search.

Thanks to GenAI, more companies than ever are deploying ML workflows. These often start as “ChatGPT wrappers,” but teams that think strategically can use early usage data to kick-start their own data flywheel, improving the user experience and building a durable advantage over potential competitors.

OpenPipe provides a fully managed workflow that fast-moving organizations are already using to automate their data flywheel. Here are the key components:

• Data collection: the OpenPipe SDK lets you automatically record your existing calls to OpenAI (and later your fine-tuned model) by changing just a few lines of code.

• Data refinement: OpenPipe lets you filter and relabel data either by hand or with LLM assistance to dramatically improve dataset quality.

• Fine tuning: OpenPipe implements an extremely streamlined fine-tuning interface for both open and closed-source models, with strong heuristics to choose the best hyperparameters but also the ability for power users to override those values.

• Evaluations: we have well-calibrated “LLM as judge” evaluations to quickly build confidence in your fine-tune’s capabilities.

• Ongoing monitoring: once your model is live in production, OpenPipe monitors its inputs and outputs using criteria you specify. This helps detect “data drift” immediately, and automatically identifies edge cases from the input domain that your model isn’t well calibrated on.

• Continuous retraining: as the monitor surfaces areas where your model performs poorly, we automatically feed those back into the training set with corrected outputs and re-train your model. This is where the data flywheel comes full circle—your model gets stronger as more of these edge cases are identified and accounted for.

Own your weights

Additionally, when you train an open-source model like Mistral 7B, Mixtral or Llama 2 through OpenPipe, you own the weights. You’re free to host them anywhere and are not dependent on us or any other company to continue providing your service.

Try us; you’ll like what you find!

It only takes a few minutes to get started with OpenPipe. You can begin gathering data immediately to fine-tune your own models and build a data flywheel—improving your users’ experience and creating a durable competitive advantage.

And finally, if this message resonates with you deeply, we’re always looking for a few extraordinary engineers to join our team in Bellevue, WA. If you think that’s you, we’d love to talk.