Artificial intelligence is going mainstream with real-world applications all around us. Enabling this shift are many startups innovating behind the scenes to improve the way AI and machine learning models are built, run, scaled, and monitored. These startups are receiving funding and being acquired by the dozens. In this article, we look at four innovative AI startups that have either been funded or acquired recently.
Cnvrg was recently acquired by Intel for an undisclosed amount. It was last valued at $17 million. Cnvrg calls itself the “operating system for machine learning and AI.” It delivers an end-to-end service for machine learning and artificial intelligence management. By this, they mean that organizations can bring multiple teams into their unified platform, build ML models, track and monitor the performance of these models, and scale them in production.
Cnvrg runs these models on container-based infrastructure or on hardware run on Nvidia machines. The container-based infrastructure is powered by Kubernetes and is highly scalable and fault-tolerant.
Cnvrg includes MLOps dashboards, which report on metrics like server and compute utilization, memory utilization, and view available capacity.
Machine learning workloads can take time to be processed because of the large volumes of data at hand. Data caching is a way of quickening this. Cnvrg provides a ready-made data caching feature that stores data within the same compute cluster and makes it instantly available for any team with access to the data.
For collaboration, Cnvrg includes interactive workspaces that let teams collaborate on ML notebooks whether they’re created from Jupyter, RStudio, or VSCode. Teams can reuse previously created notebooks and quicken their work. This would greatly improve productivity.
By thinking of every step from start to finish, Cnvrg is indeed an end-to-end ML platform. With its Intel acquisition, it now has the benefit of Intel’s worldwide customer base. This can be a big boost to the organization as it expands its reach.
Intel is on a buying spree in this space, acquiring SigOpt in the same month as Cnvrg.
SigOpt is all about optimizing your already-created ML models. SigOpt doesn’t intend to be an end-to-end solution but rather wants to be the best and most advanced AI optimization solution on the market.
SigOpt highlights the fact that it works with pretty much any data platform, including TensorFlow, PyTorch, R, AWS SageMaker, Keras, MATLAB, scikit-learn, mxnet, and more.
The first step to using SigOpt is to run initial experiments that can reproduce any data model. An experiment is aimed at improving any particular metric of a model. The experiments are very flexible, allowing you to run simple, quick runs, or complex runs with many observations and high parallelization of tasks. Experiments are run in YAML files, which are easy to store, reuse, and can be as simple or as complex as needed.
After running initial experiments, you move to the important part, which is optimization. In this step, SigOpt helps you find the parameters that result in the best metric for your model. This is done in repeated iterations until you reach an optimal accuracy. SigOpt’s optimization options include multimetric and multitask optimization and conditional parameters. At the advanced level, SigOpt also offering multimetric experiments, the ability to add prior beliefs, parameter transformation, and more. Upcoming new features include conditional experiments and better visualizations. SigOpt is making it clear they will stop at nothing to deliver the best ML model optimization solution.
The goal of all this model tuning and optimization is to make data scientists more productive. It is difficult to hire data scientists, and organizations want the ones they already have to be doing as much as they can with the available time and resources.
Along with Cnvrg, Intel could offer SigOpt as a great complementary solution for building state-of-the-art artificial intelligence models. Intel would likely work toward one unified AI solution that’s built on top of its GPUs. Intel wants to power tomorrow’s AI infrastructure, and that race is just getting started.
Chooch recently raised $20 million in Series A funding. Chooch calls itself “visual AI for any device or spectrum.” It specializes in the areas of cell identification, satellite image analysis, and pretty much anything to do with visual AI.
There are lots of real-world applications for visual artificial intelligence in every sector. Here are some examples:
- Supply chain: Use satellite imagery to track the location and speed of your fleet of vehicles.
- Health care: Use IR-sensors and X-ray data for real-time visibility into medical surgeries and procedures.
- Oil and gas: Detect oil leaks or flares before they become a crisis.
- Manufacturing: Track whether employees are following safety protocols like wearing hardhats, gloves, etc.
- Airlines: Board passengers with facial recognition.
Chooch cloud connects to cloud applications and processes data from these apps. Chooch Edge AI is meant for AIoT devices and use-cases. It is installed on any IoT device and processes image data coming from those devices.
In terms of the underlying hardware, Chooch uses Nvidia GPUs processing up to 200 API calls/sec/GPU to crunch those large volumes of image data. Chooch can process images at a speed of 0.02 seconds. They claim an accuracy of 98 percent for object recognition and 99.9 percent for facial recognition.
Chooch also offers trainable data models and a handful of pre-trained models for some use-cases to help you get started faster. The data transfer between the application and Chooch is done via APIs.
With its initial funding, Chooch is all set to start its foray into visual AI.
Abacus focuses on bringing artificial intelligence to the enterprise. It has solutions for pretty much any department in a large organization. Abacus works on deep learning models and prediction APIs to help organizations put AI to use to solve real-world challenges.
Here are some use-cases for Abacus:
- Sales and marketing: Predictive lead scoring, personalized product recommendations, customer churn rate, and sales forecasting.
- Fraud and security: Spot fraudulent transactions, detect threats, and account takeovers.
- IT operations: Cloud spend alerts, security, and vulnerability alerts.
As you can tell from the use-cases, Abacus is big on predictions rather than real-time use-cases. In these cases, the strength of the service depends on the model. These models need to be trained and optimized on a continual basis. Abacus has a ready-made list of models on their website, and they are out to make predictive modeling very easy to do on a sustained basis.
Abacus also offers a service called Deconstructed, which offers a part of its cloud AI service via APIs. You can choose just the model hosting and monitoring or use their real-time ML features store. These are meant to augment organizations’ already created machine learning models.
Abacus recently raised $13 million in Series A funding, and its future is looking promising.
Artificial intelligence startups: Good reasons they are so hot
Whether you’re looking for an end-to-end solution to unify all your data processes or a laser-focused one that solves just one area you need help with — these startups can deliver. They all approach the AI challenge from different directions, but with the same intensity and innovation that this nascent space needs. Their ability to raise funds from VCs or capital from enterprise organizations is proof of their value.
Featured image: Pixabay