Terrain Analytics Uses Intel® 4th Gen Xeon CPUs To Achieve Higher Throughput with Lower Costs

Talent acquisition is now a global game. Although the benefits that it brings are well known, the challenges that it poses for effective HR decision making aren’t talked about enough.

• Where exactly is the talent that your business needs?
• How do you know they’re the right fit?
• How much should they be costing you?

Intel® Liftoff member Terrain Analytics answers questions like these by breaking down information silos and giving HR leaders sound insights into their most valuable investment: talent. Their platform allows companies to find and hire the right people – at the right price.

Leaders in Talent Intelligence

Founded in 2020 by Mark Jacobson and Riley Kinser, Terrain is a workforce analytics data platform that companies use to make better hiring decisions. It enables Terrain’s clients to zero in on the right talent, in granular detail, down to the neighborhoods where they live. The result? Better hires, in less time, and with lower costs.

Terrain Analytics Enters Intel® Liftoff Development Sprint

In addition to blazing new trails in talent intelligence, Terrain Analytics is also a premium Intel®  Liftoff member. So we were excited to welcome them to our second development sprint on Intel® Developer Cloud for AI startups. This hackathon focused on LLM (Large Language Models) powered feature enablement. The purpose of the event was to give these startups access to accelerated computing solutions like 4th Gen Intel® Xeon® Scalable processors and Intel® Data Center GPU Max Series 1550, with a view to seeing what they could achieve with next-gen LLM capabilities.

LLMs and Their Limits

Terrain’s team opted for LLMs over traditional Machine Learning (ML) models for text classification because they have a higher success rate. They originally trialed the new features using the OpenAI Ada API, and got good results. But as powerful as OpenAI is, it is frustratingly slow.

They had used OpenAI’s Ada API to build a feature that could parse job listings. After running up against cost and throughput limitations, they decided to enter the sprint with Intel Liftoff.

Their task was to use an LLM to classify text with greater accuracy than ML, and greater speed than OpenAI. This really matters, because ML-powered solutions lack the context that an LLM has, and they tend to perform poorly with non-English text. The end goal was to be able to take several features and classify them into the labels Terrain would then use.

Once that was complete, they would need to do a second round of classifications that used the first model’s labels as a feature. This would have to be done before starting the second model training/inference steps.

Here’s what they achieved:

• Fine-tuned an LLM for the use case on Intel Data Center GPU Max
• Went on to run inference on 4th Gen Xeon CPUs
• The new model exceeded their results with Ada, with lower costs and higher throughput
  
  

Using Intel® hardware, Terrain can now efficiently scale Deep Learning/Language Learning Models (DL/LLMs) without facing computational constraints. As Terrain Analytics' data scientist, Nathan Berkley, put it: “Both of the models I made had better success metrics than the ones I built with OpenAI’s Ada model, and were 15x faster to run”. Once these models are fine-tuned, it will be possible to correctly classify job titles that are currently unclassified. Terrain will then be able to map these to their other datasets and build a feature that their users have been asking for.

Intel® Liftoff: The Tech Startup Accelerator Behind the Most Groundbreaking AI Applications

Results like this showcase what can happen when you put Intel hardware in the hands of innovative thinkers finding novel solutions to business challenges. With September’s Intel® Innovation event on the horizon, expect to see our partners continue to raise the bar on AI application development. And who knows? You could be one of them. Apply to Intel® Liftoff for Startups to find out.