≡ Menu

First Look: OneClick.ai

Share

OneClick.ai is a company taking advantage of the fact that many AI problems use similar approaches to reduce the time and cost of individual AI projects. It was founded and received its initial funding in 2017, and launched the product last year. The company has a core team of 8 in the US and China with 40 active enterprise accounts supporting over 20,000 models.

OneClick.ai uses AI to build AI and so help companies get into AI more quickly and more cheaply. The intent is to get them fault-tolerant scalable APIs for custom-built AI solutions in days or even hours instead of weeks and months. They aim to automate the end-to-end development of AI solutions based on deep learning. They use meta-learning to design and evaluate millions of deep learning models to find the best ones. They are also working on capabilities to explain how those models work, to address one of the concerns of deep learning, the lack of interpretability.

The product is aimed at non-technical users with a chatbot interface to allow experts to interact with the trained models. Users can choose from public cloud, private cloud or hosted versions and software vendors have the access to an OEM version to integrate the technology into customized solutions. A wide range of AI use cases are supported, including classic predictions (weekly and monthly sales or equipment failure) to image recognition (recognize brands in shelf images to see how much shelf space they have), classification (putting complaint emails into existing categories and identifying new problems) and semantic search (find the most helpful supporting material for a fault). Several of their existing customers were already trying to use AI and have found OneClick.ai significantly quicker to get to an accurate model.

The tool is browser-based and supports multiple projects. Each project has a chatbot that can answer data science questions. Data is provided by uploading flat files that contain a learning data set – numeric, categorical, date/time, text or images. Raw data is enough but users can add domain-specific features if they have domain knowledge that a feature will likely be helpful. Users can develop classification, regression, time-series forecasts, recommendations or clustering models and target various measures of precision depending on the type of model – accuracy, mean absolute error etc.

The engine builds many models and presents the best from which the user can select the one they prefer (based on their preferred metric and the latency of the deployed model, which is calculated for each model). The engine automatically keeps 20% out for testing and uses the other 80% for training. Under the covers, the engine keeps refining the techniques it uses based on the previous training results. Once built the chatbot can answer various questions about the models such as usage tips and model comparison. Users can deploy the models as an API for real-time access with few clicks. A future update will also allow model updates and deployment through an SDK.

You can find out more here.

Share