Ubidy Reduces CV Review Time by 95% and Improves Accuracy with AI Pipeline on Amazon Bedrock

Ubidy was born in 2016 as an online, business-to-business (B2B) recruitment marketplace connecting enterprises with specialist agencies across the globe. Using proprietary artificial intelligence (AI) pipelines, Ubidy matches global candidates with highly specific job opportunities at a fraction of the cost and time required to fill such roles via traditional large-scale agencies.

Despite the advanced AI-driven tools at Ubidy, in 2023, clients began requesting more tailored candidates for niche positions. As a customer-centric organization, Ubidy sought to refine its AI pipelines to ensure the curricula vitae (CVs) submitted to clients matched their job descriptions as closely as possible.

That same year saw the rise in a large number of new use cases for large language models (LLMs). Like other software pioneers, Ubidy is acutely aware of the AI alignment problem, in which AI often delivers biased search results not aligned with operators’ values. The company’s leaders viewed the convergence of LLM proliferation and the desire to fulfill client wishes as a unique opportunity to improve digital recruiting.

Ubidy was already in the midst of migrating its infrastructure from another cloud provider to Amazon Web Services (AWS). Evan Shellshear, who had recently joined Ubidy as managing director and chief executive officer, had long been an AWS customer and advocated for the platform’s range of tooling in fields including generative artificial intelligence (generative AI).

“If you look at all the releases and activities underway in the generative AI space, it’s clear that the breadth and depth of AWS offerings in a given period of time is consistently greater than any competitor,” Shellshear says. Ubidy compared the accuracy, performance, and cost of running LLMs like Falcon and Llama 2 on AWS versus other public clouds and open-source platforms.

Data residency and privacy were also important considerations. The nature of Ubidy’s application, which uses AI to determine the suitability of job candidates, is classified as high-risk under the European Union’s AI Act. Ubidy thus had to rule out certain open-source AI platforms because they didn’t meet data privacy requirements.

“We found that we couldn’t run our models as quickly, securely, or cost-effectively on any other platforms as we could on AWS,” Shellshear recalls. Ubidy began the migration of its entire infrastructure to AWS, a move that would facilitate integration between core services such as Amazon Elastic Compute Cloud (Amazon EC2) compute instances—which it had been using to host and manually deploy its Python machine learning (ML) pipeline—and advanced ML services.

Next, Ubidy began testing Amazon SageMaker to run its AI models. It quickly determined, however, that Amazon SageMaker wasn’t the ideal fit for its use case. An AWS solutions architect suggested Ubidy try Amazon Bedrock. Unlike Amazon SageMaker, which is a time-based service that runs 24/7, Amazon Bedrock can be deployed for processing and stopped on demand, dramatically reducing costs and architecture complexity.

“The Llama 2 70B model on Amazon Bedrock emerged as the most cost-effective option based on our evaluation criteria, providing competitive pricing without compromising performance or accuracy,” says Kah Wee Oh, chief technology and product officer at Ubidy.

The company has automated unit tests that include a fixed set of CVs and job descriptions for which it knows what the matching results should be. Each time it rolls out new models, Ubidy unit-tests against these known use cases to assess output accuracy. After successfully implementing Amazon Bedrock, Ubidy introduced its revised AI pipeline for two enterprise clients: McConnell Dowell and Europ Assistance.