H2O.ai Reviews

Normally, I use H2O.ai for my machine learning tasks, and to give an example, some of the models that I've created using H2O.ai are taxi demand forecasting and a scoring model for leads. Most of my use cases are around running machine learning algorithms on the data and then producing some sort of predictions.

I have utilized the AutoML feature in H2O.ai, which is one of the very powerful features where you don't need to worry about which algorithm is best for your model. AutoML chooses the right model for your data and does the rest itself. AutoML has been a very powerful feature in many high-end tools, such as DataRobot and DataBricks, but H2O.ai has had the AutoML feature for a long time. Although I come from a data science background and prefer to evaluate each of my models myself, I've used AutoML in H2O.ai, and I think it's a great feature to have.

AutoML reduces the time and expertise needed for developing machine learning models by approximately 50 or 60%. In traditional processes, I would choose five or six models, train each model, and then assess them based on one or multiple metrics, which can be very time-consuming. With AutoML automating that process and utilizing a large repository of models fitting on the data, it immensely reduces the time spent. Prior to AutoML, data scientists usually had four or five models in mind, but now AutoML uses more models, making it clear that a data scientist versus AutoML cannot be compared, as AutoML definitely reduces the amount of time spent on model creation.

I once worked on a model for anomaly detection, using H2O.ai to predict anomalies in terms of payments made, determining whether a payment was fraudulent. Another example includes the taxi demand prediction model, which predicts how many taxis are needed at different shopping malls in Dubai. For instance, it would predict that tomorrow at 10 a.m., 100 taxis will be needed at Dubai Mall. These are examples where H2O.ai has helped organizations make decisions based on the model output.

One of the things I really appreciate about H2O.ai is the flexibility of the tool in terms of working with different languages, as I mostly use Python for my data science work, and it's very easy to build machine learning models. H2O.ai also comes with Driverless AI, where you can automate many tasks, such as connecting to your data, performing operations on the data to produce final features for your model, training the model itself, and visualizing or saving the results, all within one platform. That's what I most appreciate about it, and you can automate that, so next time your data changes, your Driverless pipeline runs again to update your predictions or model output. The flexibility and ease of use that H2O.ai provides are very good.

One improvement I would like to see in H2O.ai is regarding the integration capabilities with different data sources, as I've seen platforms like DataIQ and DataBricks offer great integration with various data sources. H2O.ai could benefit from enhanced integration with real-time versus offline data sources, as well as improvements in productionalization solutions, including better deployment options on platforms like Azure and CI/CD integration.

One of the features I'd like to see included in upcoming releases of H2O.ai pertains to the growing trend of Generative AI, with applications for LLM-based models and vector databases. I would like to see a solution similar to Azure AI Foundry, which provides the flexibility to integrate different LLMs into applications, including H2O-GPT and other models for varied applications.

I first worked with H2O.ai in 2018, and since then, the tool has evolved. Recently, for the last six months or so, I've been using it quite extensively for different tasks, so I have been working on and off with it for about seven years.

Performance issues are quite subjective, as working with large datasets can lead to performance challenges, but overall, H2O.ai has been pretty stable for me, running very well with small to medium datasets, although larger datasets can introduce performance issues as with any platform.

H2O.ai is definitely scalable, as it supports running processes using Spark with Sparkling Water, enabling parallel execution, making it suitable for both simple and compute-intensive tasks.

I haven't escalated any questions to technical support recently, as I haven't interacted with the technical team in the last six months, but I had contact with them before.

My experience with the tech team has been good, and I don't have any negative feedback to share.

Positive

I find the initial setup of H2O.ai to be pretty straightforward, without any complexities. This contrasts with platforms like DataIQ, which I find more complicated to set up.

In my current organization, we purchased H2O.ai through a vendor, as normally the solution setup involves a bidding process with third-party vendors providing and implementing the solution.

H2O.ai helps generate positive ROI by allowing users to spend less time building models while producing more output as data scientists. The two projects I mentioned before were very successful with significant ROI impacts on business KPIs.

I would rate the technical support a nine.

For organizations considering H2O.ai, my recommendations include appreciating it as a great and flexible tool for machine learning tasks without incurring the high licensing costs associated with alternatives like DataIQ, DataRobot, and DataBricks. Users should fully utilize the AutoML functionality to streamline the model production process significantly.

On a scale of 1 to 10, I would rate H2O.ai as a product a nine, as it is a very good product and solution.

I used H2O.ai on several POCs for my previous company, and it helped me find the best model.

I needed to determine which model was performing better for job portal data. At that time, H2O.ai was evolving and had open-source libraries available for Python. I started exploring that area, and rather than working on independent or individual models, H2O.ai provided better flexibility where I could examine more models and obtain results. Based on these results, I could make the next set of decisions. With the job portal data, I was initially concerned about several models, but when I examined the dataset results from H2O.ai, I received a decent analysis which helped me finalize the model.

The best features H2O.ai offers include flexibility with AutoML, which provides the best performing model along with statistics and potential errors.

It is very stable, and when I was conducting POCs, I obtained wonderful results. Many of my colleagues also began using H2O.ai, and I observed that most banking and insurance companies started using H2O.ai as their model to get good results on the AutoML side.

It truly helped me and saves significant time because it uses in-memory computation. With good processing power, it becomes very helpful. It saves time compared to examining every individual model to understand and get the desired outcome. H2O.ai provides that flexibility from a time perspective as well.

Even though H2O.ai provides the best model, there could be improvements in certain areas. For instance, when you want to work with fusion models, H2O.ai doesn't provide that kind of information.

Currently, it provides individual models as outcomes. If it could offer combinations of models, such as suggesting using XGBoost along with SVM for wonderful results, that fusion model concept would be a good option for developers. I hope the fusion model concept will be implemented soon in H2O.ai.

Regarding documentation, I faced challenges as I didn't see much information from a documentation perspective. When I was trying to learn how to train and test H2O.ai, there was limited documentation available. If they could improve in that area, it would be really beneficial.

I have more than 15 years of experience, close to two decades. I have been in the data engineering field for the last five to seven years.

It is very stable, and when I was conducting POCs, I was getting wonderful results.

H2O.ai's scalability is quite refined, accommodating large data sets efficiently, ensuring optimal performance during model training and testing.

I evaluated Azure AutoML first. It provided good flexibility, but the issue was the extra cost associated with Azure, as everything was tied to their compute cost and their cluster cost was quite high. When I discovered the H2O Python library, it helped save considerable costs, which is why I switched from Azure AutoML to H2O.ai.

It definitely saved significant time because testing 10 models previously took an extensive amount of time and added additional cost to the company. When we started using H2O.ai, it reduced both the cost and the company expenses related to cloud infrastructure. From that perspective, it truly saved substantial cost and time.

It definitely saved significant time because testing 10 models previously took an extensive amount of time and added additional cost to the company. When we started using H2O.ai, it reduced both the cost and the company expenses related to cloud infrastructure. From that perspective, it truly saved substantial cost and time.

For larger datasets, model computation or model training and testing typically takes considerable time because with individual models, you need to train and test each one. With H2O.ai, these concerns disappear. You simply provide the training and testing data, and H2O.ai handles everything else, providing the results.

It is very easy to select models, and it's a proven AutoML solution that many industries, including insurance and banking, have been using. While many people use Azure AutoML from Azure cloud services, those seeking open source solutions with good support find H2O.ai to be excellent. It's simply a Python library you can import and utilize for numerous tasks.

I would recommend it as the best AutoML model. You can explore many features, and based on the results it provides, you can choose the best model to train your data. While I'm not familiar with the cloud-based models, H2O.ai offers a Python library that you can explore and achieve very good results.

On a scale of 1-10, I rate this solution a 7.

We mostly used the solution in the domain that I'm working. We had most of the use cases around chatbots and conversational BI.

The solution was plug-and-play, meaning most of the components were handled by the solution itself rather than building them from scratch. This was useful for our banking domain, where data privacy issues are prominent.

The most valuable feature of H2O.ai is that it is plug-and-play. Most components are handled by H2O.ai itself, which reduces the need to build from scratch.

H2O.ai can improve in areas like multimodal support and prompt engineering. They are already working on updates and changes. Although I haven't explored all the new products they've added to their list, I'm confident they might have addressed some major issues.

We used H2O.ai for two and a half years.

Since Gen AI is evolving and different application teams are reaching out with their use cases, H2O.ai might not have the full capability to handle all the use cases we face.

H2O.ai has good customer service. Being a smaller vendor with a smaller customer base, their support is quite available and exceptional whenever required.

Positive

Before using OpenAI or Vertex, we were using H2O.ai itself.

The initial setup was easy, and I would rate it as nine out of ten.

The pricing of H2O.ai was similar to that of Azure. That said, it was a bit expensive.

We are considering switching to either Azure or Google due to the scalability and growing demand in Gen AI use cases.

It is important to address data privacy concerns and ensure you're choosing the right vendor that meets your use case demands. Also, you may leave my name, Kashif, but please keep the company name private.

I'd rate the solution seven out of ten.

I am a solution architect and a consultant, and I use H2O as a machine learning platform. I create ensemble models using R and H2O, tune the hyperparameters, and then deploy them.

There are various use cases for this solution. One of the ones I worked on was a trailer forecasting solution. The customer wanted to understand the preload capacity that would be needed to have on hand so that they could call upon the right sized trailers and the right packages. It was a problem of logistics where you had to determine how many trailers were required in order to ship the packages being transported, and also have them ready just in time.

The most valuable features are the machine learning tools, the support for Jupyter Notebooks, and the collaboration that allows you to share it across people.

On the topic of model training and model governance, this solution cannot handle ten or twelve models running at the same time. It becomes a problem.

I would like to see better integration with Python and data science capabilities.

I have been using H2O for a long time.

We have had no problems with stability.

You can scale it out or scale it up, but that depends on the infrastructure. 

Our customers are Fortune 500 and Fortune 2000 companies.

I have not used technical support very much. I've been solving my own problems so I haven't had a need to use them.  I did, however, meet with the H2O guys in Seattle to work on a ticket.

I have worked with similar solutions but most of them have been custom. Not a single service could provide us with so many things. For example, H2O plus Spark, or H2O plus Sparkling Water.

The initial setup is straightforward and easy.

The deployment took between two and two and a half hours, but we did not have anybody who was an expert at the time.

H2O is a good product, and I suggest that people use it.

My advice to anybody who is considering this type of solution is to consider whether they want to procure such products and use them versus building something custom. It depends on time availability. It really just exposes Spark as the next layer.

I would rate this solution a seven out of ten.

The idea is to migrate the current model's development practice to another platform. Then after, try to create a proprietary platform using R and Python. The company is interested in using an external platform in order to have an updated environment.

Still on it. The idea is to save the cost of internal development but keeping enough flexibility to choose ML techniques and performance indicators.

Still on it.

Feature engineering.

Our primary use case is machine learning.

It has enabled our work force to be more efficient.

One of the most interesting features of the product is their driverless component. The driverless component allows you to test several different algorithms along with navigating you through choosing the best algorithm. It also gives you an interpretability capability which allows you to have some understanding of what's inside the algorithm and why it's behaving a certain way, making sure you are not bias towards the outcome.

The interpretability module has room for improvement. Also, it needs to improve its ability to integrate with other systems, like SageMaker, and the overall integration capability.

I would like more support for scalability and deep learning. Right now, they are very strong in supervise and supervise learning, but not in deep learning. I'd like to see them be more well-rounded, where they have support for deep learning, but I'm not sure that is their business model.

In terms of the stress we put on it, it is still in the very early days for us to actually take it through its phases.

It does appear to scale. We have very large use cases. The product scales as advertised. 

They have excellent tech support.

It was fairly easy to set up, then get up and running.

It was already selected. I don't know what process the company went through.

Do your due diligence, making sure with your use cases, this is the right product for you. 

Directionally, they are headed in the right place. They're also putting a lot of muscle behind it, but they're very focused in one area. Supervised on supervised learning is the market that they're going after. If that's their strategy, then they'll get some part of the market, but they'll leave the other part of the market behind.

We use just the AWS version of the product.

It integrates well with our notebooks. It also integrates well with our homegrown tool sets.

We use it for building models with large amounts of data.

We are using it for prototype projects. We have not deployed it.

The ease of use in connecting to our cluster machines.

I would like to see more features related to deployment.

As this was just for a prototype, we did not stress the product too much.

For the use case that I had, I did not run into any scaling issues. Therefore, I worked well for scalability.

I didn't run into any issues, as the application was very clear. So, I did not contact technical support.

It deals well with its core functionality. The product is definitely worth looking at, as it is one of the upcoming products where you can build large models for use cases.

I am using the on-premise version.

Our primary use case is for data science. Some of our data scientists use it pretty heavily to build models.

One example, we are able to automate life insurance. We have to underwrite policies. When somebody applies for a policy, we take their blood, then assign them a risk: substandard, standard, preferred, etc. Depending on this, we price our products. Usually the process is that you take the blood, then it goes to a lab and we get the lab results back, then an underwriter takes a look at the lab results. This is usually done in a two week time frame to get a rating. We were able to build models to automate all of this, and now, it happens in real-time. Somebody can apply online and get issued a policy right away.

It is helpful, intuitive, and easy to use. The learning curve is not too steep.

The model management features could be improved.

We haven't put a lot of stress on it.

The size of the environment for my database is probably about 900TB. 

So far, the product has been good from a scalability prospective.

I would rate the technical support as an eight out of ten.

The integration and configuration were good. I would rate them as an eight out of ten.

We have seen significant ROI where we were able to use the product in certain key projects and could automate a lot of processes. We were even able to reduce staff.

We looked at Amazon SageMaker on AWS. 

This product still was open source at that point, then we did get proprietary support after that. The other products were not open source, and we couldn't really try them out beforehand to see if we liked them or not.

H2O.ai is a great product for data scientists in general. It has a lot of options and is really flexible. Also, the pricing was good.

H2O.ai works directly with a lot of our cloud data, big data environment, and Amazon RedShift environment. The big data integration was easier from a performance perspective than Amazon RedShift. That is because our big data environment is still on-premise vs RedShift, which is on the cloud, so we had to go through some struggles to get it operating with RedShift.

We also use the on-premise version.

Testing/modeling data in the initial stages of approaching a machine-learning problem. Environment: Laptops running Ubuntu 16.04/Python 3.

AutoML helps in hands-free initial evaluations of efficiency/accuracy of ML algorithms; with training input data.

It needs a drag and drop GUI like KNIME, for easy access to and visibility of workflows.

We currently use H2O for real-time predictive analytics for fraud prevention. We have a Java-based feature engineering pipeline that attaches to POJO objects obtained from H2O. We run the whole pipeline on lightweight VMs and get under 1ms latency for each real-time transaction scoring.

We previously needed a four-machine Spark cluster to be able to train an ML model using tens of millions of transactions, and hours of time during the modeling phase. Currently, same training can now be done on an old MacBook pro with 8GB RAM within few minutes.

• Fast training
• Memory-efficient DataFrame manipulation
• Well-documented, easy-to-use algorithms
• Ability to integrate with enterprise Java apps (through POJO/MOJO) 

These are the main reasons why we switched from Spark to H2O.

Referring to bullet-3 as well, H2O DataFrame manipulation capabilities are too primitive.

Yes, we ran into a few bugs and opened JIRA tickets with reproducible test cases. We found workarounds for the problems ourselves.

No issues with scalability, it works smoothly.

We use the open-source/community branch and get support through forum discussions.

We used to developing on Scala + Spark ML. We switched, at least in part, due to reasons mentioned in the Valuable Features section of this review.

Initial setup is very easy through pip, JAR download, or R install.packages.

Currently, we do not purchase enterprise support.

We have experience with pretty much everything available; hence, the switch was an informed decision and natural.

We rate it at eight out of 10. It is very fast, light-weight, well-documented, and low-maintenance. The reasons it is not rated 10 are, it lacks the data manipulation capabilities of R and Pandas DataFrames. We would kill for dplyr offloading H2O.