Prediction Guard

🔥 Hot take: you shouldn't care about the specific AI model (ChatGPT, BLOOM, Stable Diffusion, etc.) you are using. It'll be obsolete in a matter of weeks.

So... how can you possibly keep up with the state-of-the-art and find the best model for your domain? There are 100's of thousands of models available on sites like Hugging Face, Replicate, and NGC. How do you know you are getting the best performance? How can you avoid implementation overhead when switching models and ensure that inferences are reliable?

I've been struggling to answer these questions for some time myself. There's really no good answer... until now! I built Prediction Guard so I could stop worrying about model acronyms and buggy implementations. I wanted reliable, future proof inferences via a single API call, and now I have it!

Working on some of the details, but working towards launching soon (https://www.predictionguard.com/) with features similar to:

• Automatic model selection:
  • Provide some examples of your domain specific input/ output for tasks including sentiment analysis, question answering, image captioning, speech recognition, and many more.
  • Automatically find the best model for your examples.
  • Evaluate hyped state-of-the-art models (Whisper, BLOOM, ViT, etc.) alongside battle tested classics (BERT, ResNet, etc.).
• Easy and flexible integration
  • Once automatically selected, models are available immediately via an accessible Python client and REST API.
  • Regardless of the underlying models used, the API is consistent and you don't have to worry about the implementation details.
  • Focus on how AI can solve your problems and delight your customers.
• Reliability and configurability
  • Prediction Guard has access to 100's of models for a wide range of tasks, so if your prediction fails, it will automatically fallback to the next best model.
  • You can configure the model selection process to your needs by prioritizing inference time or accuracy.
• Future proof AI endpoints
  • Prediction guard continuosly evaluates the latest models using your examples.
  • This way you can stop worrying about keeping up with the state-of-the-art and focus on your business.
  • Create your AI endpoint, and let Prediction Guard do the hard work of monitoring for better models and updating the underlying model for performance and reliability.

Have any of you experienced the pain points I've highlighted above? Of the above features/ ideas, should the value proposition focus on ease-of-use, reliability, risk avoidance, future proofing, something else? Do any competitors come to mind?