Company case
Indeed: Fine-tuning
What the published sources confirm
Indeed used OpenAI GPT models to add a natural-language explanation of why a job is recommended to a job seeker ("Invite to Apply"), then fine-tuned a smaller model to produce similar explanations with fewer tokens so it could scale.
Business problem and AI concept
- Problem
- Job recommendations without a clear reason were less persuasive and harder to trust.
- Concept
- Prompted baseline, then fine-tuning a smaller model for cost and latency, validated by A/B testing.
Reported results
OpenAI's case study reports a 20% increase in started applications and a 13% uplift in downstream success in testing.
OpenAI case study, checked 2026-10-30. Caveat: Company-reported A/B results; method details are not public.
Suggested workflow
Written for learning, not the company's system. This flow shows how you could build something similar. It is not a description of Indeed's private architecture.
Step 1: Define what a useful, honest explanation contains
Step 2: Collect consented job and profile examples
Step 3: Build a prompted baseline and label good outputs
Step 4: Fine-tune a smaller model on approved examples
Step 5: Compare quality, cost and latency on a held-out set
Step 6: A/B test with people before full rollout
Suggested technologies and alternatives
| Layer | Suggested | Alternatives |
|---|---|---|
| Fine-tuning | Hugging Face TRL + PEFT (LoRA) | Provider fine-tuning API |
| Tracking | MLflow | Weights & Biases |
| Serving | vLLM | Managed endpoint |
| Evaluation | Blind pairwise human rating | Rubric-based model grading, calibrated |
Small example
- Input
- Profile: 3 years React, wants remote. Job: Frontend engineer, remote, React + TypeScript.
- Output
- "Matches your 3 years of React experience and your preference for remote work."
Rebuild it yourself
Compare a prompted model with a LoRA-tuned small model for job-match explanations on synthetic profiles.
Definition of done
- Train/test split by profile, no overlap
- Report quality, tokens and latency for both
- Explanations never mention protected attributes
- Written decision on whether fine-tuning was worth it
Security and privacy risks
- Personal data in training sets
- Biased or discriminatory explanations
- Explanations that sound confident but are wrong
Sources
Listed by title and publisher. This guide contains no external links; search for the title if you want to read the original.
- Indeed case study. OpenAI. Vendor case study.
- How Indeed uses AI to provide better matching context for job seekers. Indeed. Company article.