putilov_denis - stock.adobe.com

Resume company builds non-biased generative AI feature

WorkStory used startup Armilla AI's AutoAlign platform. The platform tunes open source large language models and acts as a guardrail for non-open source ones.

At the start of 2023, WorkStory, Inc. wanted to take its flagship product VisualCV to the next level while keeping it free from AI bias.

The Vancouver-based company helps organizations manage multiple resumes. However, with the rise of generative AI, WorkStory wanted to upgrade VisualCV with generative AI capabilities. In its search for an AI vendor, the resume-building company came across Armilla, vendor of a responsible AI platform.

"This is the first time our company is building these capabilities into our product," said WorkStory's CEO and co-founder, Jade Bourelle. "We wanted a firm that had experience as well as technology to partner with us."

Armilla and AutoAlign

Since 2019, Armilla AI has been helping enterprises test AI model applications for specific capabilities. With the rise of generative AI, the Toronto-based vendor introduced AutoAlign on June 22.

AutoAlign is a web-based AI platform that helps remove gender, racial and other biases from large language models (LLMs) such as GPT-3 and Google Bard. It does this by finetuning open source models such as Meta Llama to remove biases or acting as a safety guardrail for models that can't be finetuned, such as GPT-3.5 or GPT-4. By guard railing non-open source products, enterprises can keep the models from answering prompts in a harmful way.

During a demonstration for TechTarget Editorial, Armilla's CEO and cofounder Dan Adamson showed how an open model was tricked into providing instructions for creating a dangerous substance such as napalm. Instead, for enterprises that use AutoAlign as a guardrail, the platform will prevent a model from producing such information.

"It's quite powerful because it doesn't require all the training sets; it can come up with those training sets and the evaluation sets and do its own tuning," Adamson said. "You can take a look at those results and evaluate it or take a look at the training sets and adjust them. But the end result is that you have a model that conforms to all of these alignment controls."

AutoAlign lets enterprises tune the LLM for bias and enables them to run and compare different models against a set of controls to determine which better suits their goal or purpose. For example, some customers started finetuning one model but ran the same finetuning test against an open-source model like Llama or RedPajama and found better results, Adamson said.

"It's very powerful to make sure you're using the right model," he said.

Screenshot of AutoAlign's platform not answering questions about Napalm.
Armilla AI's AutoAlign can keep hackers from jailbreaking an LLM and refuse to answer prompts that tricks base models into giving information on making dangerous substances.

WorkStory and AutoAlign

The ability to finetune LLMs for a specific purpose made AutoAlign appealing to a resume-building company like WorkStory. The firm gained access to the beta version of the product and tested it to build its generative AI capability into VisualCV.

"With that is also the added value [for us] to put guardrails in place," Bourelle said. "One of them being removing bias and being able to restrict things that the user shouldn't be doing."

For example, because WorkStory helps different organizations update their employees' resumes, the resume generated by the system mustn't reflect poorly on the employee.

"If you're creating automated content, and somebody puts in a description of the project that they worked on, and AI is creating that, you want to make sure that the summary of that project doesn't provide information that you wouldn't want to have," Bourelle added.

Overtuned or undertuned

This is the first time our company is building these capabilities into our product. We wanted a firm that had experience as well as technology to partner with us.
Jade BourelleCEO, WorkStory, INC

While Bourelle said the platform hasn't failed to provide whatever WorkStory wants, Adamson noted there had been times when customers have either overturned or undertuned a model, bringing about undesired results.

During his demonstration, he showed a model overtuned for gender.

Adamson put in a prompt asking an LLM to show CEOs. The original model showed all males, while the finetuned models showed all female because instead of being gender-neutral, it was tuned to be skewed to the opposite end of the spectrum.

"That's part of the power of the system but it means you have to be really careful," Adamson said. "You can also not apply things in the right way, meaning you might not put in the right alignment control or miss something."

However, despite these hiccups, Bourelle said he expects the technology to continue to evolve.

Currently, WorkStory is determining a cost model with Armilla that will depend on the resume company's volume usage. The new and improved VisualCV with generated AI capabilities is expected to be released in the fall.

Esther Ajao is a news writer covering artificial intelligence software and systems.

Dig Deeper on Enterprise applications of AI

Business Analytics
Data Management