I've got a working beta! Prediction Guard

I've got my private beta working for Prediction Guard (https://www.predictionguard.com/)! Here's the idea:

PROBLEM: How can you possibly keep up with state-of-the-art AI models? There are 100k+ models available on sites like Hugging Face. 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?

SOLUTION: Stop worrying about which AI model you should use. Just create some examples of the behavior you want, have Prediction Guard search through 100's of state-of-the-art and proven models to find one that best matches that behavior, and start getting reliable and optimized output via a simple API!

Workflow:

1. Create a few (or many) domain or use case specific examples of the type of input/output you need from a model. Many common tasks are already supported including sentiment analysis, speech-to-text, machine translation, image classification, question answering, etc...
2. Create a "proxy" endpoint. This serves as your gateway to reliable, optimized predictions. Prediction Guard evaluates all the latest, greatest AI models for you. An inference endpoint is exposed for the best one and kept up-to-date even as new models are released!

That's it! You can immediately and easily access reliable, optimized inferences via a simple API. No need to worry about the underlying model implementations or even their crazy acronyms (GPT-3.5, ViT, etc.). Just start building!

Here's how that would look with the Python client for Prediction Guard (although the same functionality is available via REST API):

import predictionguard as pg

# Initialize a Prediction Guard client.
client = pg.Client()

# Create some examples illustrating the kind of predictions you
# want to make (domain/ use case specific).
examples = [
  {
    "input": {
      "phrase": "I'm so excited about Prediction Guard. It's gonna be awesome!",
    },
    "output": {
      "sentiment": "POS"
    }
  },
  {
    "input": {
      "phrase": "AI dev without Prediction Guard sucks. It's really terrible."
    },
    "output": {
      "sentiment": "NEG"
    }
  }
]

# Create a prediction "proxy." This proxy will save your examples, evaluate
# SOTA models to find the best one for your use case, and expose the best model
# at an endpoint corresponding to the proxy.
client.create_proxy(task='sentiment', name='test-client-sentiment3', 
                    examples=examples)

# Now your ready to start getting reliable, future proof predictions. No fuss!
client.predict(name='test-client-sentiment2', data={
    "phrase": "Isn't this great! Sign up for the waitlist to get in on the fun."
})

See more detailed feature descriptions and sign up for the waitlist here: https://www.predictionguard.com/

I would love feedback on this idea, the functionality, and how you might envision using it in your own projects! Does the pitch above resonate? What doesn't make sense? Where might be you see the value in this functionality? Would you use the above via Python, REST API, JavaScript, other? Really appreciate this community and any input you might have!