For more than seven years, UpCodes has been making the complicated world of building codes easier to understand for people like industry professionals and homeowners. Its platform includes a searchable database covering regulations in all states and features like a “spellcheck” that flags code errors. Today, the startup is announcing a new AI-based tool that will make navigating the world of building codes even more streamlined. Called Copilot and built on ChatGPT-4, Copilot serves as a research assistant, answering complicated code questions and annotating responses with links to relevant sections of code.
UpCodes also announced it has closed a Series A of $3.5 million, intended for hiring as it continues developing Copilot and adding more AI-based functions to its platform. Along with UpCodes’ previous funding, including a pre-Series A announced in March 2021, this brings its total raised to $7.6 million.
The latest round was led by Building Ventures, a VC firm focused on construction and real estate tech. Other participants include PlanGrid’s co-founders, CapitalX and Bragiel Bros.
UpCodes now has more than 650,000 monthly active users and has served over 100 million page views. Since TechCrunch last covered UpCodes in March 2021, it’s grown quite a lot. Scott Reynolds, its co-founder and CEO, said the startup’s team has doubled, its revenue has quadrupled and its product offerings have expanded to include more user segments. The building code it covers has also increased from less than two million to more than five million sections hosted, and it now offers coverage for all U.S. states and major cities.
Before launching Copilot, UpCodes was focused on building its database of code, often digitizing regulations that were only available in physical reference books, and making them easier to look up. In addition to over five million code sections, it also hosts 160,000 local amendments. Codes are constantly changing, so UpCodes updates about 7,000 each month on average.
Its database is searchable and has other tools designed to make code compliance easier, like its code check feature, but even using those takes a lot of time because of how complex regulations are. Copilot is meant to dramatically simplify the code research process.
An example of how Copilot answers questions about building codes.
Reynolds gave some examples of the questions Copilot can answer:
- Calculate the travel distance for a specific occupancy in a building, or the maximum distance someone can travel in case of an emergency, to an exit (an example of Copilot’s response to this type of query is included in the graphic above).
- Elaborate on the context behind a code section to help understand the meaning.
- Find related or more stringent code sections for other codes such as building, fire and mechanical code.
- Generate a checklist for residential deck regulations, with relevant code sections linked.
Copilot answers those questions and helps users by citing the code sections it pulls information from, so they can review the actual code themselves.
“We’ve always leaned heavily into education and helping users understand the underlying context,” Reynolds said. UpCodes plans to add more explanatory content to Copilot that will help users further understand code, beyond publicly available content.
Notably, laws hosted by UpCodes have contributed to the top 0.01% of training data for AIs like ChatGPT and Google Bard. Building professionals often turn to those tools to get their questions answered, Reynolds said. They were trained on the C4 dataset from Common Crawl, which pulled directly from UpCodes’ website (basic access to its code database is free).
“There’s an increase in focus on data quality for training LLMs,” he explained. “UpCodes has an extensive library of high-quality construction law, which is ideal for an LLM to train on as we’re the only source online for many of these laws. It may have been less of a conscious decision to include building laws and more a result of their algorithm identifying quality data relevant across a broad range of topics, including construction laws.”
But Reynolds added that the crawl only includes snapshots in time, so their models are likely operating from out-of-date code since regulations are constantly changing.
This gives Copilot an advantage, since it draws on UpCodes’ constantly updated database. On top of that, codes also vary widely from jurisdiction to jurisdiction. Reynolds said UpCodes spent years building the infrastructure to keep its code database up to date.
One of the most important challenges facing any AI project is limiting hallucinations. Reynolds said one of the biggest steps UpCodes has taken is “fencing” Copilot into applicable codes for a project. This means it parses over five million code sections, including 160,000 amendments, based on a user’s location and permit year. It also collects additional data like type of building to make sure Copilot only uses relevant codes.
Copilot has an internal system that fine-tunes responses specific to construction law, and passes each query through multiple layers of analysis to better understand and parse questions. It focuses on jurisdiction-specific codes, to account for variances and because building professionals often have to be familiar with several jurisdictions at once. Once a question is asked, Copilot gives answers with context and direct references so users can see how it arrived at an answer.
Users who are subscribed to an UpCodes paid plan get to ask Copilot three questions. If they want unlimited access, they can upgrade to UpCodes Professional or add it to their Enterprise plan.
UpCodes’ Series A will be used to hire for its engineering department, as well as every department that contributes to Copilot. It plans to expand its code library and resources so Copilot can produce more sophisticated answers and add new features, like project management, Reynolds said.
The startup’s team has known Building Ventures since it was founded. “We thought they’d be the perfect partner for our Series A,” he added, explaining that the firm is made up of former operators in the construction industry, including former founders. Its portfolio and LP base serve as a valuable resource for guidance when UpCodes needs to validate ideas.
“The construction industry can be opaque, a world unto itself, so having investors and partners steeped in it for decades is helpful,” Reynolds said.
In a statement about Building Ventures’ investment into UpCodes, partner Allen Preger said, “By unifying and maintaining all building codes in an AI-powered platform, UpCodes is transforming code compliance for the Built Environment. We are thrilled to be leading their Series A investment.”
UpCodes launches Copilot, an AI-based research assistant for building codes by Catherine Shu originally published on TechCrunch