Need autonomous driving training data? ›

Case Study: Init.ai

“The expertise in generating instructions has been critical. The time saved is immense.”

Keith Brisson,
CEO at Init.ai

The Company

Init.ai is a venture-backed technology company founded in 2015 and based in NYC and San Francisco. The company offers businesses and developers a platform through which to build, train, and deploy intelligent conversation-based applications and conversational-understanding tools using artificial intelligence (AI), natural language processing (NLP), and messaging.

As the Init.ai website explains, “Our developer platform can be used to create standalone, platform-agnostic apps that require no downloads. It can also be used to add artificial intelligence and chat into customer apps. We offer a comprehensive solution that handles messaging, machine learning, and business logic, including integration with third-party APIs.”

Init.ai has set out to create tooling that “enables companies to understand and respond using natural language,” explained Keith Brisson, Init.ai CEO. But their sights are on an even bigger picture: they’re developing a full-on “conversational understanding system” with the goal of providing a more holistic, conversation-based vision of dialogue through technology. They are not interested in processing single messages only, but instead creating AI that considers and comprehends whole conversations to extract intent and meaning. Init.ai aims to empower ecommerce companies, assist support teams, and more with this advanced technology.

Chatbots, as they’re colloquially known, are essentially product #1 for the young company—but they’re supercharged chatbots, given that they leverage Init.ai’s proprietary NLP and AI capabilities. Plus, bots built through the Init.ai are platform-agnostic; they integrate with all of the popular messaging channels and services. Future products will augment human agents as they serve customers and also interact through multi-user chats to enable frictionless suggestions and event monitoring.

The Challenge

To teach a system to understand conversations, you need data—specifically, conversations—to train it. The Init.ai team considered setting up their own chatbots and having people chat with them, but knew the resulting conversations would be restricted in scope due to existing bots’ limited natural language understanding—the very problem they were trying to solve. They tried coming up with sample conversations on their own and staging dialogue with colleagues, but it was obvious that wouldn’t scale. Those conversations were once again limited, this time by the worldviews and vocabularies of a very small group of chatters.

Next they set up “have a conversation with me” tasks on a crowdsourcing platform, but struggled to design the task and write the instructions in a way that helped people “get it” and that yielded high-quality, usable training conversations quickly and efficiently.

The team soon realized they needed a new approach. Though the crowdsourcing option seemed like a solution to the scale problem, it ultimately wasn’t a good solution. Brisson explained: “They cost us a fair amount of money and took a lot of time for our staff.”

The Move to Mighty AI

Mighty AI’s history of producing training data for machine learning for other companies was something that initially intrigued Init.ai about the TDaaS solution. “We recognized the expertise there,” explained Brisson. “We’re not experts in training data collection.”

After a few meetings to learn about Mighty AI’s solution, approach, and team, and to explain their goals and challenges, Init.ai signed on and Mighty AI got to work creating conversation-generation tasks.

The Solution

Mighty AI designed tasks in which members of our tasking community chatted with each other in role-playing scenarios, pretending to be both customers and support agents or company representatives. Our understanding of Init.ai’s goals combined with our expertise in running training data tasks and and managing taskers informed how we designed the task and what we wrote in the task instructions. This was an important piece of the puzzle: ”The expertise in generating the instructions and how to make these tasks is what has been most critical,” shared Brisson.

Once conversations were created and filtered for quality, our community also classified individual “utterances”—single phrases within the larger conversations—as assertions, questions, filler speech, etc.

The Results

Mighty AI and Init.ai have iterated on the project together, consistently adjusting for higher quality and increased efficiency, though results have been encouraging from the beginning: “We were actually pleasantly surprised by the first batch [of data] we got back,” Brisson commented. “The variety was higher quality than we expected.”

As for ease and efficiency? In comparison to the crowdsourcing solution they’d experimented with previously, Brisson said, “The time saved is immense.”

Will you be our next success story?
Let’s discuss how Mighty AI can help.