Sculpt

@gobi_dasu Thank you, and thanks for the question. Here's how we look at it: Tools can be divided into: 1) whether they focus on in-house annotators with domain expertise (Us, Prodigy) vs crowdsourced labeling orgs that manage huge teams of annotators (Scale, Figure Eight, Mechanical Turk, etc). In-house teams usually give you an advantage in speed, higher accuracy for more complex/subjective tasks, and security-sensitive tasks. 2) intelligent (use ML to improve the whole process) vs UI that randomly goes over all the data points. Intelligent systems usually reach your desired accuracy faster and cheaper, although it's a delicate balance because it could insert bias or lower the quality of the dataset. We currently see ourselves as an in-house, intelligent data labeling tool. Also here are some of our key advantages in more detail: 1) We've worked hard to optimize the whole process to be very intelligent, time efficient, and customizable. For example: Sculpt uses data augmentation techniques to automatically label thousands of examples, gives active feedback so you stop when you reach the performance you need, lets you examine and edit the features that your model uses to make predictions, etc. 2) We don't charge per annotation, or per API call. We allow you to download the model from Sculpt as a pickled file and deploy it anywhere, and to download the datasets so you can train your own models. 3) Finally, tools like Prodigy also use ML such as active learning to speed up the training process. We agree with their philosophy but we think it's missing key components like active feedback, quality assurance, data augmentation, and giving users control over how the model makes predictions.

@gobi_dasu if you have your own views of the space I'd love to hear them out

@abraham_starosta1 great answer

@sid_grover Thanks for asking! 1) yes no problem, we have done that with other customers before 2) yes, we would need more details but yes we can give guidance on this and help figure this out

@_jayhack_ happy to see Mira doing great!

@nissim_darwiche Yeah! you can download the dataset and train your own models (Bert, XLNet, etc)

@avitlkn Hey Shai! no we don't charge for predictions. We give 3 options: 1) you can download your dataset and train your own model 2) you can download a pickled model file from Sculpt and deploy it wherever you want (for no extra charge) 3) or we can deploy the model for you and give you an API (for which we would charge). The charge would depend on the volume, but would be low price (still figuring that out)

@mariano_sorgente thank you! you don't need to know AI, I'd say the only thing you need to understand are common model performance metrics like precision and recall