Evolving with AI to Transform Consumer Energy Experiences

“Is spending thousands of dollars on solar panels the right decision for me?”

These are the kinds of questions we answer at WattBuy.

To that end, we’ve spent the last five years honing our machine learning models to provide personalized energy insights and product recommendations for households across the US. But one of our users' biggest pain points is the lack of access to and understanding of this information as it pertains to solar. Enter Sunny, our AI-powered solar advisor chatbot.

Sunny integrates with our tools and data to let customers have an open-ended and relevant conversation about their solar options. We built Sunny by using OpenAI’s LLM and integrating it with our own energy APIs to provide the chatbot with valuable data and capabilities that allow it to answer our customers’ very specific questions, empowering them to feel confident in their decisions while avoiding interactions with pushy solar sales people.

At WattBuy, we have a culture of continuous innovation, so when DeepSeek was released, our Solutions Architect Idris Jaffer set out to build a quick proof of concept to test how easily we could switch over to DeepSeek from OpenAI. Turns out, he was very successful.

We are currently using the DeepSeek R1 Distill Llama 70B model on AWS, which leverages its expansive 70 billion parameters. This “distilled” model is the result of training a smaller model to mimic a larger model’s predictions and internal reasoning, yielding similar performance with reduced computational cost. We found that DeepSeek was basically a slot-in replacement for OpenAI, in terms of standard chatbot capabilities, and even with the deeper API integrations we had already built. We changed the name of the model and the URL it was pointing to, and everything else just worked.

And what about the cost? The majority of our cost for Sunny comes from OpenAI. With published API pricing, we found that DeepSeek costs roughly 20% of the cost of OpenAI. Not 20% less than OpenAI. 20% of the cost of OpenAI. This is huge, with DeepSeek R1 offering a significant savings opportunity over the pricing for OpenAI GPT-4o. One caveat here is that we plan on using DeepSeek as a downloaded, open source instance, not using the DeepSeek API. We can’t compare self-hosting costs directly, since OpenAI doesn’t have an open source option, but are hoping the cost savings extend to most versions of DeepSeek.

How about performance? A full comparison of the performance of various LLMs is beyond the scope of this proof of concept, but benchmarks between GPT-4o (our current platform) and DeepSeek R1 indicate that R1 typically surpasses 4o in most categories. Anecdotally, though, it seems that DeepSeek is less eloquent than our current platform, OpenAI GPT-4o. As a “reasoning” model, however, DeepSeek R1 can explain its reasoning, and justify its calculations (similar to OpenAI’s o1 model). This is important for Sunny, since many of our users have high distrust of the information provided by the solar industry, yet struggle to understand the complexity behind solar numbers. We'll continue experimenting with additional tuning and optimizations to see if we can bridge the gap between R1 and GPT-4o for Sunny.

Finally, we are excited to provide increased privacy and security for our users. Since DeepSeek R1 is open source and downloadable, we don’t need to connect to someone else’s server to use it. Hosting our own model, we don’t have to worry about the possibility of our customers' data being shared outside our systems, or our customers' interactions being used to train another model. We haven’t really considered if DeepSeek’s app or APIs are privacy risks, but we love the practice of offering downloadable, open source models, allowing companies to keep sensitive data in-house and private. It may be harder to compare the price of hosting our model against APIs, but this seems worth it to ensure our customers’ privacy.

AI platforms are rapidly changing and quickly providing new tools for agile startups willing to experiment. We're thrilled at the relative ease and low cost with which we were able to address a pain point blocking people from getting cheaper and cleaner energy. If we collectively use AI in this manner—to help people better their lives—imagine the world we can create.

If you're using AI in innovative ways, we'd love to hear about it! Please reach out.