Users Dig Deep with WattBuy’s Gen-AI Innovations
At WattBuy, we have spent the last 5 years building out a suite of Machine Learning models for personalized energy insights. These models take in a user’s address, then use that address to assemble details related to the structure of their home; the solar potential of their roof; and the weather, climate, and cost of electricity in their area. All of this information is used to generate a wide variety of insights, from predicted energy usage to product recommendations. But one of our users' biggest pain points is the lack of access to and understanding of this information to make a confident decision on rooftop solar. Enter Sunny, our AI-powered solar advisor chatbot.
AI Background
Over the last few years, we’ve seen a proliferation of Generative AI tools, commonly in the form of conversational assistants (AKA chatbots). At their core, these chatbots are able to have conversations with the user and pull on widely available information that may have been present in their original training data. In order to further tailor these general assistants to specific use cases, finetuning and retrieval-augmented generation (RAG) became commonplace in the industry.
Finetuning allows developers to add additional information into the training of the model. This is great for changing the “core” of how the model operates. For example, a model that needs to read highly technical documents might be finetuned on similar documents so that it has a better ability to understand what it is reading.
On the other hand, RAG sits on top of the model and allows a generalized model to access specific information right as the user needs it. For example, an AI agent may be able to retrieve information about a specific ecommerce product when requested by a user.
Problem
Finetuning and RAG are great for a wide variety of use cases, but don’t quite get to the full extent of what we at WattBuy envisioned for Sunny. While both of these solutions can access static information, we needed to be able to personalize the feedback or even generate new data in real time. This would allow Sunny to learn new information during the conversation and then actually use that information to generate completely new insights.
At the same time, we had a valuable set of ML-driven insights that we had developed over five years. These tools were being used behind the scenes to drive user experiences like product recommendations and bill estimates, but we knew that they could be used for so much more with the right interface.
Solution
With these two considerations in mind, we set out to combine our longstanding expertise in ML-driven energy modeling with this new wave of advancements in Generative AI. The result is Sunny, an AI conversational chatbot that is empowered to run advanced energy modeling and cost scenarios for users in order to inform conversations about available solar options.
To accomplish this, we started with the foundations of a standard RAG approach. We use an AI Routing Agent to analyze user intent with each prompt, then compare that intent against a list of available external sources. If relevant, the external data is gathered and included with the user’s request to the main Conversational AI Agent. For some queries, this looks exactly like a traditional RAG approach. For example, a user asking for a list of solar incentives in their area might trigger our Routing Agent to query our incentives database and pass that information along down the chain.
However, where our approach really starts to shine is when a user asks something like “how much will my electric bill be in October?”, or even “how long will it take me to break even on these solar panels?”. Because we’ve built this Routing Agent ourselves, we were able to create custom abilities that go far beyond the simple data retrieval found in off the shelf products.
Our Routing Agent sees the intent in those questions and, rather than querying a database of static information, it runs Machine Learning models in order to inform our response. In the case of the solar break-even point, the Routing Agent might use our Electricity Usage ML Model, Solar Cost ML Model, and Solar Production ML Model. These would be combined with traditional RAG methodology to retrieve electricity cost information, available tax incentives, and anything else that the model considers relevant. All of this data is then passed on to our main Conversational Agent, which distills it into an informed response to the user’s specific question.
Because our chatbot is able to run all of these models in real time, the user is then able to continue to dig deeper into individual data points, or even provide clarifying information to tweak the response. For example, the user may provide new information about their vacation habits; the chat bot is able to interpret this, make an educated guess on its impact, and adapt on the fly.
Databricks
Databricks has been instrumental in this process from the start. Even before the recent explosion of Generative AI technologies, Databricks’ Data Intelligence Platform provided the foundation for building AI & ML use cases on top of our existing datasets. We store all of our data in the Delta format in our Lakehouse, and use Mlflow’s wide range of ML Ops capabilities to train, test, iterate, and publish our models. These tools are the building blocks of the AI System we have today.
Once we decided to jump into the world of conversational AI, not only were we already well-positioned to build something extraordinary, but we had the partners to help us. We built something unique in the AI (and clean energy!) world and it was essential to have the Databricks team to collaborate with and push us forward. Together, we built something exciting and valuable.
Not only that, Databricks has now released an interface for others to develop similar projects via the AI Agent Tools Preview. This is part of Databricks’ broader Mosaic AI platform, in which customers can create, iterate on, and deploy agent systems on their own data. We’re really excited to streamline our process with this new development and to see how others begin using the same techniques.
"As we march toward global clean energy goals, WattBuy has built a multi-faceted AI-powered system that makes clean energy more accessible than ever before", said Julien Debard, Director of Energy and Utilities at Databricks. "This system was made possible through a combination of WattBuy's deep industry expertise, foundational machine learning models, and the end-to-end AI capabilities of Databricks' Data Intelligence Platform. We look forward to leveraging data and AI to build more clean energy solutions together."
Results & Conclusion
Does our unique integration between our ML models and a chatbot provide our customers with more helpful information than with a traditional AI Chatbot? Absolutely. We find that customers presented with the results of our ML models typically have an average conversation length about four times longer than if they don’t receive any of the ML model results. They ask more follow up questions, and they interact with the data more, finding the value of continuing the conversation with this marriage of ML based models and a chatbot.
How do we know? Many of those customers decide to move forward with a formal solar inquiry. We find that about one in ten customers who interact with at least one ML API call also eventually request a solar inquiry form, by asking more questions indicating they want to take the next step to sign up for solar. Users also provide more positive feedback (a quick thumbs up), about 5.2 times more, on conversations with at least one ML API call, compared to conversations that don’t include an ML API call.
WattBuy was created to give customers actionable information about their energy usage. Over several years, we’ve built a huge number of tools and calculators for customers to estimate their electricity usage, their costs, and their solar costs and production potential. Now, with Sunny, we can provide customers with a completely new, fully integrated interface to explore all of our energy tools, so they can make the best informed decisions for their energy future.
Written by Mason Hollis at WattBuy with Sohaib Vora at Databricks