All that glitters is not gold ✨
by Orlando Becerra and Rob Wright
The Use Case
Apploi is revolutionizing healthcare hiring with its all-in-one platform designed to recruit, onboard, and schedule top healthcare professionals. Serving over 9,000 healthcare organizations across the United States, we address one of the industry’s biggest challenges: delivering exceptional care amidst a candidate shortage.
In collaboration with a partner, we provide job postings to candidates in both English and Spanish. Currently, we use Google Translate to translate every job in our system, ensuring they are accessible to Spanish-speaking audiences. However, this method has become increasingly expensive due to the high volume of daily translations.
The Challenge
With the escalating costs, our primary objective was to find a more cost-effective solution without compromising the quality of translations. We decided to explore various tools that appeared promising, akin to “gold,” in addressing our needs.
Exploring Alternative Tools
Meta’s SeamlessM4T
The first tool we evaluated was Meta’s SeamlessM4T, a cutting-edge text-to-text translation service supporting nearly 100 languages. While it’s free, it requires hosting, which we managed. Initial tests showed that it translated quickly and accurately, a real “golden” feature for us. However, it fell short when handling job descriptions with embedded HTML, a crucial requirement for us to maintain the original formatting. As a result, this tool wasn’t the gold we hoped for.
OpenAI’s GPT-3
Next, we turned to OpenAI’s GPT-3. This tool promised significant cost savings — up to 90% — by replacing Google Translate in our codebase. The transition was straightforward, and we thoroughly documented the trade-off study and conducted a initial implementation. Despite the potential savings, the average job description took 90 seconds to translate with GPT-3 compared to just one second with Google Translate. Given our volume of many, many thousands of jobs per month, even with optimizations this would lead to significant delays and higher infrastructure costs, making it another false gold for our needs.
The Turnaround
Apploi is a company that has grown rapidly. When we first added the job translation feature four years ago, we had a lot fewer job postings and our business requirements were much different. With all the hype and excitement, we were likely pulled into the Generative AI black hole. (And honestly it was fun!) But this project forced us to ask a fundamental question. What are the business requirements today? We discovered that not every job needed to be translated; only those relevant to specific companies or job postings required translation. This insight led us to a much simpler and more cost-effective solution: continue to use Google Translate but implement a feature flag to toggle translations on or off for specific job postings or companies.
Conclusion
By focusing on the precise needs of our business, we significantly reduced costs and achieved our goals in an unexpected way. This experience taught us that what appears to be gold might not always be the best solution, and sometimes, simplicity is the true treasure.