We built Pyq because it took too long to build features that were powered by AI at our previous jobs. A lot of people want to get started using AI, but struggle because of the difficulties involved in managing infrastructure, finding the right model and learning how to call it. There are many interesting and useful models in places like Github or Hugging Face, as well as specific applications of popular models like OpenAI’s GPT, but they require a decent amount of work/knowledge to get working in your app.

The first issue is determining if and how your problem can be solved with AI. This generally involves experimenting with different models and (more recently) prompts, followed by potentially fine-tuning your model, at which point you’ll have to repeat this process with datasets. Then you move onto the set of challenges posed by getting that model deployed in production, including messing around with Docker, cloud infrastructure etc. This process can take weeks or even months. We aim to make it easy to match a problem to an AI solution and get it working in your application quickly.

Aman was leading a product team at a startup and was told that an already-built AI model would take an additional 3 weeks to bring to production. The only solution was to hire an engineer to do this and potentially pay for an enterprise MLOps platform on top of that. Simultaneously, Emily at Microsoft found herself asking the Azure team directly for help to hook up a model into the HoloLens application she was working on. The ensuing frustration resulted in our first principle: bringing an AI model to production should take minutes, not weeks!

Infrastructure is only one part of the problem. With all of the new possibilities afforded by modern AI models, it can be difficult to understand what business applications they can be used for. We decided to apply our knowledge of building AI-powered products to finding practical use cases that are easy for any developer to understand, even if they don’t have any AI knowledge.

We identify use cases of various AI models and provide straightforward APIs tailored to those use cases. We use both open-source models and popular providers such as OpenAI. This allows for easy and fast integration into apps. Rather than starting with the model, experimenting to see if it can actually do what you want it to, learning about deployment and serving, developers can just make a POST call to start using AI.

We serve our models with FastAPI, containerize them, and then deploy them to our GKE clusters. Depending on the model, we choose different machines - some require GPUs, most are decent on CPU. We take models up or down based on usage, so we have cold starts unless otherwise specified by customers. We expose access to the model via a POST call through our cloud app. We track inputs and outputs, as we expect that people will become interested in fine tuning models based on their past usage.

Pyq is not meant for AI experts or specialists, but for people who are building features which are powered by AI. We have a curated list of models that are good at specific tasks and are inexpensive to use. Some have been used thousands of times already!

Deploying your own model with us is also a very straightforward process and can usually be done within an hour. For those requiring low latency and high volume, we also offer a high performance API at additional cost.

Shortly after the launch of Chat GPT, we created a GPT Detector (https://www.gpt-detector.com, also available via API through our website) in collaboration with another YC company. This got a surprising amount of traction due to the virality of ChatGPT itself. Building the entire website took less than a day - we fine-tuned an existing text classification model, deployed it on Pyq and our partner integrated it with their front-end. It has been used 10,000+ times since then, and has been quite performant and inexpensive.

We have seen several other applications created in a similar way using Pyq. These include document OCR apps, chatbots, stock image generators and more.

We have a prepaid, usage-based pricing model. Every model has a “spot price” - the cost of 1 second of compute. This is available on each model’s page in our ‘Zoo.’ If you deploy your own model, we will give you your initial price manually and adjust it up or down over time depending on your needs.

We also provide $10 of free computing credit upon signup. This is enough to experiment with all of our models and, for some of them, enough to run a few hundred or even a thousand inferences. We add more credits on an ad-hoc basis, so feel free to email us at team[at]pyqai.com describing what you’re working on and we’ll do our best to accommodate you!

We are so excited to show this product. Our hope is that it helps you bring a project to life, finish that feature you’ve been working on, or just gives you ideas for what to build next. Please weigh in and tell us what you think!