2023 was a whirlwind for software startup HData. Over the course of the year, the 27-person company pulled off a Series A raise, an expansion into generative artificial intelligence, and a pilot with energy giant Southern Company.
Those steps set the scene for the months ahead, during which the company plans to expand its suite of energy regulatory tools — and to prove that utilities using its platform ultimately make more progress on renewables adoption than others.
- The top line: HData made its debut in the energy regulatory data market in 2020, using artificial intelligence to pull numbers from regulatory PDFs: everything from utility expenses to hourly load. It wasn’t until last year, though, that the company started to deploy generative AI, using natural language processing to comb and analyze documents. That move will have a huge influence on the tools and partnerships that the company pursues from here, CEO Hudson Hollister told Latitude Media.
- The nuts and bolts: Rather than building its own AI models, HData uses a combination of third-party large language models as the foundation of both its platform and its generative AI capabilities. These include OpenAI’s ChatGPT, Google’s Gemini, and Anthropic’s Claude.
- The current take: The energy industry is still grappling with how and when to put generative AI to work, and how complicated it might be to do so, Hollister said. “We know about large language models now, we know it’s possible to train a model for a specific purpose. The part that I think is not yet fully understood is that oftentimes, you don’t need to,” he added. “What really matters is not so much building a model, as taking the existing ones, which are really good, and feeding them properly.”
HData uses a process called retrieval augmented generation to give the models it subscribes to additional information — pulled from regulatory documents — to inform their responses. That information, Hollister explained, is broken down into smaller chunks based on where in a document they’re pulled from, and stored in the company’s database.
Then, when a user asks a complex question, such as why the California Public Utilities Commission approved an electric vehicle plan for the utility PG&E, HData’s technology identifies which chunks are relevant to that query and feeds them to the LLMs, which then generate a response.
“Using this retrieval augmented generation technology, we can get accurate answers — and we do not have the hallucination that would occur if you were just using one of those famous large language models by itself,” Hollister explained. “We also can do citations back to the source, and we’ve built that architecture [to be] specialized for energy regulation, because our customers tell us what kind of answers they need.”
There is unlikely to be a future in which HData builds and trains its own models, Hollister added.
“I think we’ll be building skills, [and] I’ll bet most of the skills won’t require their own models either,” he said. “I think they’ll require feeding an existing model in a certain way.”
Well-fed models at work
The first utility to put HData’s new generative AI capabilities to work was Southern Company, in a pilot program that the pair launched last summer. The program, which is still underway, is aimed at assessing the feasibility and benefits of the platform for departments such as regulatory accounting and finance.
In the months since the pilot began, Southern Company has run around 1,400 projects on the HData platform in ten different offices, Hollister said, including those of its subsidiaries Alabama Power, Georgia Power, and Mississippi Power.
“One of the things that’s most exciting for me is seeing more of the different possible applications,” he said. “There’s dozens of different offices inside a large utility that need to get questions answered out of this corpus of regulatory information.”
Those projects have ranged from tracking and comparing capacity and actual generation of power plants, to charting the finances of individual transmission wires. Doing these things effectively involves tracking down valuable data that’s buried deep inside regulatory filings, Hollister said. As a result of that pilot, HData will roll out a brand new tool on the platform in late April.
Ultimately, the company’s contribution to the load growth conundrum facing the power sector today is to smooth regulatory approvals — and ideally boost renewables penetration. Matching innovation and generation to the needs of the industry starts by breaking down the “unholy table” of years worth of hourly load data, Hollister said, as well as speeding up the interconnection queue, which he believes can be done by analyzing paperwork faster.
“We need to work better with the numbers and the words in order to understand exactly what we’re dealing with,” he added.
