Indeed

Solution

Indeed combined its long-standing job-matching AI with OpenAI GPT models to improve the language and context behind Invite to Apply recommendations. The team first used the Chat Completions API and few-shot prompting to generate richer explanations for why a candidate matched a specific job, then realized the prompt-heavy approach consumed too many tokens at Indeed's scale. In response, Indeed and OpenAI fine-tuned a smaller GPT model using GPT-4 for data augmentation, an annotation operation with labeled outputs and ground truths, and content guidelines that taught the model to explain relevance economically. Dedicated instances let the engineering team deploy the fine-tuned model at very high volume, personalizing recommendations to millions more job seekers. The result is a recommendation system that still depends on Indeed's proprietary marketplace data and matching logic, but now adds a natural-language rationale that improves both application starts and downstream hiring success.

Data flow

Indeed combines resume, profile, and qualification data with job-posting context, then passes the matched candidate-job pair to the fine-tuned GPT model for a natural-language explanation. The generated rationale is sent through Invite to Apply at scale on dedicated instances and measured through A/B tests against downstream hiring outcomes.